Merge "Add unit test for VP9 denoiser."
diff --git a/test/convolve_test.cc b/test/convolve_test.cc
index d1e7f1d..99bdf82 100644
--- a/test/convolve_test.cc
+++ b/test/convolve_test.cc
@@ -21,6 +21,9 @@
#include "vpx_ports/mem.h"
namespace {
+
+static const int kMaxDimension = 64;
+
typedef void (*ConvolveFunc)(const uint8_t *src, ptrdiff_t src_stride,
uint8_t *dst, ptrdiff_t dst_stride,
const int16_t *filter_x, int filter_x_stride,
@@ -30,9 +33,10 @@
struct ConvolveFunctions {
ConvolveFunctions(ConvolveFunc h8, ConvolveFunc h8_avg,
ConvolveFunc v8, ConvolveFunc v8_avg,
- ConvolveFunc hv8, ConvolveFunc hv8_avg)
+ ConvolveFunc hv8, ConvolveFunc hv8_avg,
+ int bd)
: h8_(h8), v8_(v8), hv8_(hv8), h8_avg_(h8_avg), v8_avg_(v8_avg),
- hv8_avg_(hv8_avg) {}
+ hv8_avg_(hv8_avg), use_high_bd_(bd) {}
ConvolveFunc h8_;
ConvolveFunc v8_;
@@ -40,6 +44,7 @@
ConvolveFunc h8_avg_;
ConvolveFunc v8_avg_;
ConvolveFunc hv8_avg_;
+ int use_high_bd_; // 0 if high bitdepth not used, else the actual bit depth.
};
typedef std::tr1::tuple<int, int, const ConvolveFunctions *> ConvolveParam;
@@ -68,71 +73,66 @@
const int kInterp_Extend = 4;
const unsigned int intermediate_height =
(kInterp_Extend - 1) + output_height + kInterp_Extend;
+ unsigned int i, j;
- /* Size of intermediate_buffer is max_intermediate_height * filter_max_width,
- * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
- * + kInterp_Extend
- * = 3 + 16 + 4
- * = 23
- * and filter_max_width = 16
- */
- uint8_t intermediate_buffer[71 * 64];
+ // Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+ // where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+ // + kInterp_Extend
+ // = 3 + 16 + 4
+ // = 23
+ // and filter_max_width = 16
+ //
+ uint8_t intermediate_buffer[71 * kMaxDimension];
const int intermediate_next_stride = 1 - intermediate_height * output_width;
// Horizontal pass (src -> transposed intermediate).
- {
- uint8_t *output_ptr = intermediate_buffer;
- const int src_next_row_stride = src_stride - output_width;
- unsigned int i, j;
- src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
- for (i = 0; i < intermediate_height; ++i) {
- for (j = 0; j < output_width; ++j) {
- // Apply filter...
- const int temp = (src_ptr[0] * HFilter[0]) +
- (src_ptr[1] * HFilter[1]) +
- (src_ptr[2] * HFilter[2]) +
- (src_ptr[3] * HFilter[3]) +
- (src_ptr[4] * HFilter[4]) +
- (src_ptr[5] * HFilter[5]) +
- (src_ptr[6] * HFilter[6]) +
- (src_ptr[7] * HFilter[7]) +
- (VP9_FILTER_WEIGHT >> 1); // Rounding
+ uint8_t *output_ptr = intermediate_buffer;
+ const int src_next_row_stride = src_stride - output_width;
+ src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+ for (i = 0; i < intermediate_height; ++i) {
+ for (j = 0; j < output_width; ++j) {
+ // Apply filter...
+ const int temp = (src_ptr[0] * HFilter[0]) +
+ (src_ptr[1] * HFilter[1]) +
+ (src_ptr[2] * HFilter[2]) +
+ (src_ptr[3] * HFilter[3]) +
+ (src_ptr[4] * HFilter[4]) +
+ (src_ptr[5] * HFilter[5]) +
+ (src_ptr[6] * HFilter[6]) +
+ (src_ptr[7] * HFilter[7]) +
+ (VP9_FILTER_WEIGHT >> 1); // Rounding
- // Normalize back to 0-255...
- *output_ptr = clip_pixel(temp >> VP9_FILTER_SHIFT);
- ++src_ptr;
- output_ptr += intermediate_height;
- }
- src_ptr += src_next_row_stride;
- output_ptr += intermediate_next_stride;
+ // Normalize back to 0-255...
+ *output_ptr = clip_pixel(temp >> VP9_FILTER_SHIFT);
+ ++src_ptr;
+ output_ptr += intermediate_height;
}
+ src_ptr += src_next_row_stride;
+ output_ptr += intermediate_next_stride;
}
// Vertical pass (transposed intermediate -> dst).
- {
- uint8_t *src_ptr = intermediate_buffer;
- const int dst_next_row_stride = dst_stride - output_width;
- unsigned int i, j;
- for (i = 0; i < output_height; ++i) {
- for (j = 0; j < output_width; ++j) {
- // Apply filter...
- const int temp = (src_ptr[0] * VFilter[0]) +
- (src_ptr[1] * VFilter[1]) +
- (src_ptr[2] * VFilter[2]) +
- (src_ptr[3] * VFilter[3]) +
- (src_ptr[4] * VFilter[4]) +
- (src_ptr[5] * VFilter[5]) +
- (src_ptr[6] * VFilter[6]) +
- (src_ptr[7] * VFilter[7]) +
- (VP9_FILTER_WEIGHT >> 1); // Rounding
+ src_ptr = intermediate_buffer;
+ const int dst_next_row_stride = dst_stride - output_width;
+ for (i = 0; i < output_height; ++i) {
+ for (j = 0; j < output_width; ++j) {
+ // Apply filter...
+ const int temp = (src_ptr[0] * VFilter[0]) +
+ (src_ptr[1] * VFilter[1]) +
+ (src_ptr[2] * VFilter[2]) +
+ (src_ptr[3] * VFilter[3]) +
+ (src_ptr[4] * VFilter[4]) +
+ (src_ptr[5] * VFilter[5]) +
+ (src_ptr[6] * VFilter[6]) +
+ (src_ptr[7] * VFilter[7]) +
+ (VP9_FILTER_WEIGHT >> 1); // Rounding
- // Normalize back to 0-255...
- *dst_ptr++ = clip_pixel(temp >> VP9_FILTER_SHIFT);
- src_ptr += intermediate_height;
- }
- src_ptr += intermediate_next_stride;
- dst_ptr += dst_next_row_stride;
+ // Normalize back to 0-255...
+ *dst_ptr++ = clip_pixel(temp >> VP9_FILTER_SHIFT);
+ src_ptr += intermediate_height;
}
+ src_ptr += intermediate_next_stride;
+ dst_ptr += dst_next_row_stride;
}
}
@@ -159,16 +159,137 @@
unsigned int dst_stride,
unsigned int output_width,
unsigned int output_height) {
- uint8_t tmp[64 * 64];
+ uint8_t tmp[kMaxDimension * kMaxDimension];
- assert(output_width <= 64);
- assert(output_height <= 64);
+ assert(output_width <= kMaxDimension);
+ assert(output_height <= kMaxDimension);
filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
output_width, output_height);
block2d_average_c(tmp, 64, dst_ptr, dst_stride,
output_width, output_height);
}
+#if CONFIG_VP9_HIGHBITDEPTH
+void high_filter_block2d_8_c(const uint16_t *src_ptr,
+ const unsigned int src_stride,
+ const int16_t *HFilter,
+ const int16_t *VFilter,
+ uint16_t *dst_ptr,
+ unsigned int dst_stride,
+ unsigned int output_width,
+ unsigned int output_height,
+ int bd) {
+ // Between passes, we use an intermediate buffer whose height is extended to
+ // have enough horizontally filtered values as input for the vertical pass.
+ // This buffer is allocated to be big enough for the largest block type we
+ // support.
+ const int kInterp_Extend = 4;
+ const unsigned int intermediate_height =
+ (kInterp_Extend - 1) + output_height + kInterp_Extend;
+
+ /* Size of intermediate_buffer is max_intermediate_height * filter_max_width,
+ * where max_intermediate_height = (kInterp_Extend - 1) + filter_max_height
+ * + kInterp_Extend
+ * = 3 + 16 + 4
+ * = 23
+ * and filter_max_width = 16
+ */
+ uint16_t intermediate_buffer[71 * kMaxDimension];
+ const int intermediate_next_stride = 1 - intermediate_height * output_width;
+
+ // Horizontal pass (src -> transposed intermediate).
+ {
+ uint16_t *output_ptr = intermediate_buffer;
+ const int src_next_row_stride = src_stride - output_width;
+ unsigned int i, j;
+ src_ptr -= (kInterp_Extend - 1) * src_stride + (kInterp_Extend - 1);
+ for (i = 0; i < intermediate_height; ++i) {
+ for (j = 0; j < output_width; ++j) {
+ // Apply filter...
+ const int temp = (src_ptr[0] * HFilter[0]) +
+ (src_ptr[1] * HFilter[1]) +
+ (src_ptr[2] * HFilter[2]) +
+ (src_ptr[3] * HFilter[3]) +
+ (src_ptr[4] * HFilter[4]) +
+ (src_ptr[5] * HFilter[5]) +
+ (src_ptr[6] * HFilter[6]) +
+ (src_ptr[7] * HFilter[7]) +
+ (VP9_FILTER_WEIGHT >> 1); // Rounding
+
+ // Normalize back to 0-255...
+ *output_ptr = clip_pixel_high(temp >> VP9_FILTER_SHIFT, bd);
+ ++src_ptr;
+ output_ptr += intermediate_height;
+ }
+ src_ptr += src_next_row_stride;
+ output_ptr += intermediate_next_stride;
+ }
+ }
+
+ // Vertical pass (transposed intermediate -> dst).
+ {
+ uint16_t *src_ptr = intermediate_buffer;
+ const int dst_next_row_stride = dst_stride - output_width;
+ unsigned int i, j;
+ for (i = 0; i < output_height; ++i) {
+ for (j = 0; j < output_width; ++j) {
+ // Apply filter...
+ const int temp = (src_ptr[0] * VFilter[0]) +
+ (src_ptr[1] * VFilter[1]) +
+ (src_ptr[2] * VFilter[2]) +
+ (src_ptr[3] * VFilter[3]) +
+ (src_ptr[4] * VFilter[4]) +
+ (src_ptr[5] * VFilter[5]) +
+ (src_ptr[6] * VFilter[6]) +
+ (src_ptr[7] * VFilter[7]) +
+ (VP9_FILTER_WEIGHT >> 1); // Rounding
+
+ // Normalize back to 0-255...
+ *dst_ptr++ = clip_pixel_high(temp >> VP9_FILTER_SHIFT, bd);
+ src_ptr += intermediate_height;
+ }
+ src_ptr += intermediate_next_stride;
+ dst_ptr += dst_next_row_stride;
+ }
+ }
+}
+
+void high_block2d_average_c(uint16_t *src,
+ unsigned int src_stride,
+ uint16_t *output_ptr,
+ unsigned int output_stride,
+ unsigned int output_width,
+ unsigned int output_height,
+ int bd) {
+ unsigned int i, j;
+ for (i = 0; i < output_height; ++i) {
+ for (j = 0; j < output_width; ++j) {
+ output_ptr[j] = (output_ptr[j] + src[i * src_stride + j] + 1) >> 1;
+ }
+ output_ptr += output_stride;
+ }
+}
+
+void high_filter_average_block2d_8_c(const uint16_t *src_ptr,
+ const unsigned int src_stride,
+ const int16_t *HFilter,
+ const int16_t *VFilter,
+ uint16_t *dst_ptr,
+ unsigned int dst_stride,
+ unsigned int output_width,
+ unsigned int output_height,
+ int bd) {
+ uint16_t tmp[kMaxDimension * kMaxDimension];
+
+ assert(output_width <= kMaxDimension);
+ assert(output_height <= kMaxDimension);
+ high_filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter, tmp, 64,
+ output_width, output_height, bd);
+ high_block2d_average_c(tmp, 64, dst_ptr, dst_stride,
+ output_width, output_height, bd);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
class ConvolveTest : public ::testing::TestWithParam<ConvolveParam> {
public:
static void SetUpTestCase() {
@@ -177,6 +298,13 @@
vpx_memalign(kDataAlignment, kInputBufferSize + 1)) + 1;
output_ = reinterpret_cast<uint8_t*>(
vpx_memalign(kDataAlignment, kOutputBufferSize));
+#if CONFIG_VP9_HIGHBITDEPTH
+ input16_ = reinterpret_cast<uint16_t*>(
+ vpx_memalign(kDataAlignment,
+ (kInputBufferSize + 1) * sizeof(uint16_t))) + 1;
+ output16_ = reinterpret_cast<uint16_t*>(
+ vpx_memalign(kDataAlignment, (kOutputBufferSize) * sizeof(uint16_t)));
+#endif
}
static void TearDownTestCase() {
@@ -184,6 +312,12 @@
input_ = NULL;
vpx_free(output_);
output_ = NULL;
+#if CONFIG_VP9_HIGHBITDEPTH
+ vpx_free(input16_ - 1);
+ input16_ = NULL;
+ vpx_free(output16_);
+ output16_ = NULL;
+#endif
}
protected:
@@ -191,7 +325,6 @@
static const int kOuterBlockSize = 256;
static const int kInputStride = kOuterBlockSize;
static const int kOutputStride = kOuterBlockSize;
- static const int kMaxDimension = 64;
static const int kInputBufferSize = kOuterBlockSize * kOuterBlockSize;
static const int kOutputBufferSize = kOuterBlockSize * kOuterBlockSize;
@@ -212,6 +345,12 @@
virtual void SetUp() {
UUT_ = GET_PARAM(2);
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ != 0)
+ mask_ = (1 << UUT_->use_high_bd_) - 1;
+ else
+ mask_ = 255;
+#endif
/* Set up guard blocks for an inner block centered in the outer block */
for (int i = 0; i < kOutputBufferSize; ++i) {
if (IsIndexInBorder(i))
@@ -222,15 +361,25 @@
::libvpx_test::ACMRandom prng;
for (int i = 0; i < kInputBufferSize; ++i) {
- if (i & 1)
+ if (i & 1) {
input_[i] = 255;
- else
+#if CONFIG_VP9_HIGHBITDEPTH
+ input16_[i] = mask_;
+#endif
+ } else {
input_[i] = prng.Rand8Extremes();
+#if CONFIG_VP9_HIGHBITDEPTH
+ input16_[i] = prng.Rand16() & mask_;
+#endif
+ }
}
}
void SetConstantInput(int value) {
memset(input_, value, kInputBufferSize);
+#if CONFIG_VP9_HIGHBITDEPTH
+ vpx_memset16(input16_, value, kInputBufferSize);
+#endif
}
void CheckGuardBlocks() {
@@ -240,20 +389,123 @@
}
}
- uint8_t* input() const {
+ uint8_t *input() const {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
+ } else {
+ return CONVERT_TO_BYTEPTR(input16_ + BorderTop() * kOuterBlockSize +
+ BorderLeft());
+ }
+#else
return input_ + BorderTop() * kOuterBlockSize + BorderLeft();
+#endif
}
- uint8_t* output() const {
+ uint8_t *output() const {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
+ } else {
+ return CONVERT_TO_BYTEPTR(output16_ + BorderTop() * kOuterBlockSize +
+ BorderLeft());
+ }
+#else
return output_ + BorderTop() * kOuterBlockSize + BorderLeft();
+#endif
+ }
+
+ uint16_t lookup(uint8_t *list, int index) const {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ return list[index];
+ } else {
+ return CONVERT_TO_SHORTPTR(list)[index];
+ }
+#else
+ return list[index];
+#endif
+ }
+
+ void assign_val(uint8_t *list, int index, uint16_t val) const {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ list[index] = (uint8_t) val;
+ } else {
+ CONVERT_TO_SHORTPTR(list)[index] = val;
+ }
+#else
+ list[index] = (uint8_t) val;
+#endif
+ }
+
+ void wrapper_filter_average_block2d_8_c(const uint8_t *src_ptr,
+ const unsigned int src_stride,
+ const int16_t *HFilter,
+ const int16_t *VFilter,
+ uint8_t *dst_ptr,
+ unsigned int dst_stride,
+ unsigned int output_width,
+ unsigned int output_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+ dst_ptr, dst_stride, output_width,
+ output_height);
+ } else {
+ high_filter_average_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
+ HFilter, VFilter,
+ CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
+ output_width, output_height,
+ UUT_->use_high_bd_);
+ }
+#else
+ filter_average_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+ dst_ptr, dst_stride, output_width,
+ output_height);
+#endif
+ }
+
+ void wrapper_filter_block2d_8_c(const uint8_t *src_ptr,
+ const unsigned int src_stride,
+ const int16_t *HFilter,
+ const int16_t *VFilter,
+ uint8_t *dst_ptr,
+ unsigned int dst_stride,
+ unsigned int output_width,
+ unsigned int output_height) {
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0) {
+ filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+ dst_ptr, dst_stride, output_width, output_height);
+ } else {
+ high_filter_block2d_8_c(CONVERT_TO_SHORTPTR(src_ptr), src_stride,
+ HFilter, VFilter,
+ CONVERT_TO_SHORTPTR(dst_ptr), dst_stride,
+ output_width, output_height, UUT_->use_high_bd_);
+ }
+#else
+ filter_block2d_8_c(src_ptr, src_stride, HFilter, VFilter,
+ dst_ptr, dst_stride, output_width, output_height);
+#endif
}
const ConvolveFunctions* UUT_;
static uint8_t* input_;
static uint8_t* output_;
+#if CONFIG_VP9_HIGHBITDEPTH
+ static uint16_t* input16_;
+ static uint16_t* output16_;
+ int mask_;
+#endif
};
+
uint8_t* ConvolveTest::input_ = NULL;
uint8_t* ConvolveTest::output_ = NULL;
+#if CONFIG_VP9_HIGHBITDEPTH
+uint16_t* ConvolveTest::input16_ = NULL;
+uint16_t* ConvolveTest::output16_ = NULL;
+#endif
TEST_P(ConvolveTest, GuardBlocks) {
CheckGuardBlocks();
@@ -272,7 +524,8 @@
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
- ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
+ ASSERT_EQ(lookup(out, y * kOutputStride + x),
+ lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
@@ -289,7 +542,8 @@
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
- ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
+ ASSERT_EQ(lookup(out, y * kOutputStride + x),
+ lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
@@ -306,7 +560,8 @@
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
- ASSERT_EQ(out[y * kOutputStride + x], in[y * kInputStride + x])
+ ASSERT_EQ(lookup(out, y * kOutputStride + x),
+ lookup(in, y * kInputStride + x))
<< "(" << x << "," << y << ")";
}
@@ -339,8 +594,18 @@
TEST_P(ConvolveTest, MatchesReferenceSubpixelFilter) {
uint8_t* const in = input();
uint8_t* const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+ uint8_t ref8[kOutputStride * kMaxDimension];
+ uint16_t ref16[kOutputStride * kMaxDimension];
+ uint8_t* ref;
+ if (UUT_->use_high_bd_ == 0) {
+ ref = ref8;
+ } else {
+ ref = CONVERT_TO_BYTEPTR(ref16);
+ }
+#else
uint8_t ref[kOutputStride * kMaxDimension];
-
+#endif
for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
const InterpKernel *filters =
@@ -350,10 +615,10 @@
for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
- filter_block2d_8_c(in, kInputStride,
- filters[filter_x], filters[filter_y],
- ref, kOutputStride,
- Width(), Height());
+ wrapper_filter_block2d_8_c(in, kInputStride,
+ filters[filter_x], filters[filter_y],
+ ref, kOutputStride,
+ Width(), Height());
if (filters == eighttap_smooth || (filter_x && filter_y))
ASM_REGISTER_STATE_CHECK(
@@ -375,7 +640,8 @@
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
- ASSERT_EQ(ref[y * kOutputStride + x], out[y * kOutputStride + x])
+ ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+ lookup(out, y * kOutputStride + x))
<< "mismatch at (" << x << "," << y << "), "
<< "filters (" << filter_bank << ","
<< filter_x << "," << filter_y << ")";
@@ -387,16 +653,36 @@
TEST_P(ConvolveTest, MatchesReferenceAveragingSubpixelFilter) {
uint8_t* const in = input();
uint8_t* const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+ uint8_t ref8[kOutputStride * kMaxDimension];
+ uint16_t ref16[kOutputStride * kMaxDimension];
+ uint8_t* ref;
+ if (UUT_->use_high_bd_ == 0) {
+ ref = ref8;
+ } else {
+ ref = CONVERT_TO_BYTEPTR(ref16);
+ }
+#else
uint8_t ref[kOutputStride * kMaxDimension];
+#endif
// Populate ref and out with some random data
::libvpx_test::ACMRandom prng;
for (int y = 0; y < Height(); ++y) {
for (int x = 0; x < Width(); ++x) {
- const uint8_t r = prng.Rand8Extremes();
+ uint16_t r;
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0 || UUT_->use_high_bd_ == 8) {
+ r = prng.Rand8Extremes();
+ } else {
+ r = prng.Rand16() & mask_;
+ }
+#else
+ r = prng.Rand8Extremes();
+#endif
- out[y * kOutputStride + x] = r;
- ref[y * kOutputStride + x] = r;
+ assign_val(out, y * kOutputStride + x, r);
+ assign_val(ref, y * kOutputStride + x, r);
}
}
@@ -408,10 +694,10 @@
for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
- filter_average_block2d_8_c(in, kInputStride,
- filters[filter_x], filters[filter_y],
- ref, kOutputStride,
- Width(), Height());
+ wrapper_filter_average_block2d_8_c(in, kInputStride,
+ filters[filter_x], filters[filter_y],
+ ref, kOutputStride,
+ Width(), Height());
if (filters == eighttap_smooth || (filter_x && filter_y))
ASM_REGISTER_STATE_CHECK(
@@ -433,7 +719,8 @@
for (int y = 0; y < Height(); ++y)
for (int x = 0; x < Width(); ++x)
- ASSERT_EQ(ref[y * kOutputStride + x], out[y * kOutputStride + x])
+ ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+ lookup(out, y * kOutputStride + x))
<< "mismatch at (" << x << "," << y << "), "
<< "filters (" << filter_bank << ","
<< filter_x << "," << filter_y << ")";
@@ -442,6 +729,103 @@
}
}
+TEST_P(ConvolveTest, FilterExtremes) {
+ uint8_t *const in = input();
+ uint8_t *const out = output();
+#if CONFIG_VP9_HIGHBITDEPTH
+ uint8_t ref8[kOutputStride * kMaxDimension];
+ uint16_t ref16[kOutputStride * kMaxDimension];
+ uint8_t *ref;
+ if (UUT_->use_high_bd_ == 0) {
+ ref = ref8;
+ } else {
+ ref = CONVERT_TO_BYTEPTR(ref16);
+ }
+#else
+ uint8_t ref[kOutputStride * kMaxDimension];
+#endif
+
+ // Populate ref and out with some random data
+ ::libvpx_test::ACMRandom prng;
+ for (int y = 0; y < Height(); ++y) {
+ for (int x = 0; x < Width(); ++x) {
+ uint16_t r;
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (UUT_->use_high_bd_ == 0 || UUT_->use_high_bd_ == 8) {
+ r = prng.Rand8Extremes();
+ } else {
+ r = prng.Rand16() & mask_;
+ }
+#else
+ r = prng.Rand8Extremes();
+#endif
+ assign_val(out, y * kOutputStride + x, r);
+ assign_val(ref, y * kOutputStride + x, r);
+ }
+ }
+
+ for (int axis = 0; axis < 2; axis++) {
+ int seed_val = 0;
+ while (seed_val < 256) {
+ for (int y = 0; y < 8; ++y) {
+ for (int x = 0; x < 8; ++x) {
+#if CONFIG_VP9_HIGHBITDEPTH
+ assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+ ((seed_val >> (axis ? y : x)) & 1) * mask_);
+#else
+ assign_val(in, y * kOutputStride + x - SUBPEL_TAPS / 2 + 1,
+ ((seed_val >> (axis ? y : x)) & 1) * 255);
+#endif
+ if (axis) seed_val++;
+ }
+ if (axis)
+ seed_val-= 8;
+ else
+ seed_val++;
+ }
+ if (axis) seed_val += 8;
+
+ for (int filter_bank = 0; filter_bank < kNumFilterBanks; ++filter_bank) {
+ const InterpKernel *filters =
+ vp9_get_interp_kernel(static_cast<INTERP_FILTER>(filter_bank));
+ const InterpKernel *const eighttap_smooth =
+ vp9_get_interp_kernel(EIGHTTAP_SMOOTH);
+ for (int filter_x = 0; filter_x < kNumFilters; ++filter_x) {
+ for (int filter_y = 0; filter_y < kNumFilters; ++filter_y) {
+ wrapper_filter_block2d_8_c(in, kInputStride,
+ filters[filter_x], filters[filter_y],
+ ref, kOutputStride,
+ Width(), Height());
+ if (filters == eighttap_smooth || (filter_x && filter_y))
+ ASM_REGISTER_STATE_CHECK(
+ UUT_->hv8_(in, kInputStride, out, kOutputStride,
+ filters[filter_x], 16, filters[filter_y], 16,
+ Width(), Height()));
+ else if (filter_y)
+ ASM_REGISTER_STATE_CHECK(
+ UUT_->v8_(in, kInputStride, out, kOutputStride,
+ kInvalidFilter, 16, filters[filter_y], 16,
+ Width(), Height()));
+ else
+ ASM_REGISTER_STATE_CHECK(
+ UUT_->h8_(in, kInputStride, out, kOutputStride,
+ filters[filter_x], 16, kInvalidFilter, 16,
+ Width(), Height()));
+
+ for (int y = 0; y < Height(); ++y)
+ for (int x = 0; x < Width(); ++x)
+ ASSERT_EQ(lookup(ref, y * kOutputStride + x),
+ lookup(out, y * kOutputStride + x))
+ << "mismatch at (" << x << "," << y << "), "
+ << "filters (" << filter_bank << ","
+ << filter_x << "," << filter_y << ")";
+ }
+ }
+ }
+ }
+ }
+}
+
DECLARE_ALIGNED(256, const int16_t, kChangeFilters[16][8]) = {
{ 0, 0, 0, 0, 0, 0, 0, 128},
{ 0, 0, 0, 0, 0, 0, 128},
@@ -505,7 +889,8 @@
kPixelSelected + ((kInitialSubPelOffset
+ kFilterPeriodAdjust * kInputPixelStep)
>> SUBPEL_BITS);
- ASSERT_EQ(in[ref_x], out[x]) << "x == " << x << "width = " << Width();
+ ASSERT_EQ(lookup(in, ref_x), lookup(out, x))
+ << "x == " << x << "width = " << Width();
}
/* Test the vertical filter. */
@@ -520,7 +905,8 @@
kPixelSelected + ((kInitialSubPelOffset
+ kFilterPeriodAdjust * kInputPixelStep)
>> SUBPEL_BITS);
- ASSERT_EQ(in[ref_y * kInputStride], out[y * kInputStride]) << "y == " << y;
+ ASSERT_EQ(lookup(in, ref_y * kInputStride), lookup(out, y * kInputStride))
+ << "y == " << y;
}
/* Test the horizontal and vertical filters in combination. */
@@ -543,7 +929,8 @@
+ kFilterPeriodAdjustX * kInputPixelStep)
>> SUBPEL_BITS);
- ASSERT_EQ(in[ref_y * kInputStride + ref_x], out[y * kOutputStride + x])
+ ASSERT_EQ(lookup(in, ref_y * kInputStride + ref_x),
+ lookup(out, y * kOutputStride + x))
<< "x == " << x << ", y == " << y;
}
}
@@ -570,7 +957,8 @@
for (int y = 0; y < Height(); ++y) {
for (int x = 0; x < Width(); ++x) {
- ASSERT_EQ(in[y * kInputStride + x], out[y * kOutputStride + x])
+ ASSERT_EQ(lookup(in, y * kInputStride + x),
+ lookup(out, y * kOutputStride + x))
<< "x == " << x << ", y == " << y
<< ", frac == " << frac << ", step == " << step;
}
@@ -582,11 +970,479 @@
using std::tr1::make_tuple;
#if CONFIG_VP9_HIGHBITDEPTH
+#if HAVE_SSE2 && ARCH_X86_64
+void wrap_convolve8_horiz_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 8);
+}
+
+void wrap_convolve8_avg_horiz_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_vert_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_avg_vert_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 8);
+}
+
+void wrap_convolve8_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_avg_sse2_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_horiz_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_horiz_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_vert_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_vert_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_sse2_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 10);
+}
+
+void wrap_convolve8_horiz_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_avg_horiz_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_vert_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_avg_vert_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 12);
+}
+
+void wrap_convolve8_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 12);
+}
+
+void wrap_convolve8_avg_sse2_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_sse2(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 12);
+}
+#endif // HAVE_SSE2 && ARCH_X86_64
+
+void wrap_convolve8_horiz_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_avg_horiz_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_vert_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_avg_vert_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 8);
+}
+
+void wrap_convolve8_avg_c_8(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 8);
+}
+
+void wrap_convolve8_horiz_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_horiz_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 10);
+}
+
+void wrap_convolve8_vert_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_vert_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_avg_c_10(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride, w, h, 10);
+}
+
+void wrap_convolve8_horiz_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_avg_horiz_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_vert_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_avg_vert_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+void wrap_convolve8_avg_c_12(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x,
+ int filter_x_stride,
+ const int16_t *filter_y,
+ int filter_y_stride,
+ int w, int h) {
+ vp9_high_convolve8_avg_c(src, src_stride, dst, dst_stride, filter_x,
+ filter_x_stride, filter_y, filter_y_stride,
+ w, h, 12);
+}
+
+const ConvolveFunctions convolve8_c(
+ wrap_convolve8_horiz_c_8, wrap_convolve8_avg_horiz_c_8,
+ wrap_convolve8_vert_c_8, wrap_convolve8_avg_vert_c_8,
+ wrap_convolve8_c_8, wrap_convolve8_avg_c_8, 8);
+INSTANTIATE_TEST_CASE_P(C_8, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve8_c),
+ make_tuple(8, 4, &convolve8_c),
+ make_tuple(4, 8, &convolve8_c),
+ make_tuple(8, 8, &convolve8_c),
+ make_tuple(16, 8, &convolve8_c),
+ make_tuple(8, 16, &convolve8_c),
+ make_tuple(16, 16, &convolve8_c),
+ make_tuple(32, 16, &convolve8_c),
+ make_tuple(16, 32, &convolve8_c),
+ make_tuple(32, 32, &convolve8_c),
+ make_tuple(64, 32, &convolve8_c),
+ make_tuple(32, 64, &convolve8_c),
+ make_tuple(64, 64, &convolve8_c)));
+const ConvolveFunctions convolve10_c(
+ wrap_convolve8_horiz_c_10, wrap_convolve8_avg_horiz_c_10,
+ wrap_convolve8_vert_c_10, wrap_convolve8_avg_vert_c_10,
+ wrap_convolve8_c_10, wrap_convolve8_avg_c_10, 10);
+INSTANTIATE_TEST_CASE_P(C_10, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve10_c),
+ make_tuple(8, 4, &convolve10_c),
+ make_tuple(4, 8, &convolve10_c),
+ make_tuple(8, 8, &convolve10_c),
+ make_tuple(16, 8, &convolve10_c),
+ make_tuple(8, 16, &convolve10_c),
+ make_tuple(16, 16, &convolve10_c),
+ make_tuple(32, 16, &convolve10_c),
+ make_tuple(16, 32, &convolve10_c),
+ make_tuple(32, 32, &convolve10_c),
+ make_tuple(64, 32, &convolve10_c),
+ make_tuple(32, 64, &convolve10_c),
+ make_tuple(64, 64, &convolve10_c)));
+const ConvolveFunctions convolve12_c(
+ wrap_convolve8_horiz_c_12, wrap_convolve8_avg_horiz_c_12,
+ wrap_convolve8_vert_c_12, wrap_convolve8_avg_vert_c_12,
+ wrap_convolve8_c_12, wrap_convolve8_avg_c_12, 12);
+INSTANTIATE_TEST_CASE_P(C_12, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve12_c),
+ make_tuple(8, 4, &convolve12_c),
+ make_tuple(4, 8, &convolve12_c),
+ make_tuple(8, 8, &convolve12_c),
+ make_tuple(16, 8, &convolve12_c),
+ make_tuple(8, 16, &convolve12_c),
+ make_tuple(16, 16, &convolve12_c),
+ make_tuple(32, 16, &convolve12_c),
+ make_tuple(16, 32, &convolve12_c),
+ make_tuple(32, 32, &convolve12_c),
+ make_tuple(64, 32, &convolve12_c),
+ make_tuple(32, 64, &convolve12_c),
+ make_tuple(64, 64, &convolve12_c)));
+
#else
+
const ConvolveFunctions convolve8_c(
vp9_convolve8_horiz_c, vp9_convolve8_avg_horiz_c,
vp9_convolve8_vert_c, vp9_convolve8_avg_vert_c,
- vp9_convolve8_c, vp9_convolve8_avg_c);
+ vp9_convolve8_c, vp9_convolve8_avg_c, 0);
INSTANTIATE_TEST_CASE_P(C, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_c),
@@ -606,11 +1462,65 @@
#if HAVE_SSE2 && ARCH_X86_64
#if CONFIG_VP9_HIGHBITDEPTH
+const ConvolveFunctions convolve8_sse2(
+ wrap_convolve8_horiz_sse2_8, wrap_convolve8_avg_horiz_sse2_8,
+ wrap_convolve8_vert_sse2_8, wrap_convolve8_avg_vert_sse2_8,
+ wrap_convolve8_sse2_8, wrap_convolve8_avg_sse2_8, 8);
+INSTANTIATE_TEST_CASE_P(SSE2_8, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve8_sse2),
+ make_tuple(8, 4, &convolve8_sse2),
+ make_tuple(4, 8, &convolve8_sse2),
+ make_tuple(8, 8, &convolve8_sse2),
+ make_tuple(16, 8, &convolve8_sse2),
+ make_tuple(8, 16, &convolve8_sse2),
+ make_tuple(16, 16, &convolve8_sse2),
+ make_tuple(32, 16, &convolve8_sse2),
+ make_tuple(16, 32, &convolve8_sse2),
+ make_tuple(32, 32, &convolve8_sse2),
+ make_tuple(64, 32, &convolve8_sse2),
+ make_tuple(32, 64, &convolve8_sse2),
+ make_tuple(64, 64, &convolve8_sse2)));
+const ConvolveFunctions convolve10_sse2(
+ wrap_convolve8_horiz_sse2_10, wrap_convolve8_avg_horiz_sse2_10,
+ wrap_convolve8_vert_sse2_10, wrap_convolve8_avg_vert_sse2_10,
+ wrap_convolve8_sse2_10, wrap_convolve8_avg_sse2_10, 10);
+INSTANTIATE_TEST_CASE_P(SSE2_10, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve10_sse2),
+ make_tuple(8, 4, &convolve10_sse2),
+ make_tuple(4, 8, &convolve10_sse2),
+ make_tuple(8, 8, &convolve10_sse2),
+ make_tuple(16, 8, &convolve10_sse2),
+ make_tuple(8, 16, &convolve10_sse2),
+ make_tuple(16, 16, &convolve10_sse2),
+ make_tuple(32, 16, &convolve10_sse2),
+ make_tuple(16, 32, &convolve10_sse2),
+ make_tuple(32, 32, &convolve10_sse2),
+ make_tuple(64, 32, &convolve10_sse2),
+ make_tuple(32, 64, &convolve10_sse2),
+ make_tuple(64, 64, &convolve10_sse2)));
+const ConvolveFunctions convolve12_sse2(
+ wrap_convolve8_horiz_sse2_12, wrap_convolve8_avg_horiz_sse2_12,
+ wrap_convolve8_vert_sse2_12, wrap_convolve8_avg_vert_sse2_12,
+ wrap_convolve8_sse2_12, wrap_convolve8_avg_sse2_12, 12);
+INSTANTIATE_TEST_CASE_P(SSE2_12, ConvolveTest, ::testing::Values(
+ make_tuple(4, 4, &convolve12_sse2),
+ make_tuple(8, 4, &convolve12_sse2),
+ make_tuple(4, 8, &convolve12_sse2),
+ make_tuple(8, 8, &convolve12_sse2),
+ make_tuple(16, 8, &convolve12_sse2),
+ make_tuple(8, 16, &convolve12_sse2),
+ make_tuple(16, 16, &convolve12_sse2),
+ make_tuple(32, 16, &convolve12_sse2),
+ make_tuple(16, 32, &convolve12_sse2),
+ make_tuple(32, 32, &convolve12_sse2),
+ make_tuple(64, 32, &convolve12_sse2),
+ make_tuple(32, 64, &convolve12_sse2),
+ make_tuple(64, 64, &convolve12_sse2)));
#else
const ConvolveFunctions convolve8_sse2(
vp9_convolve8_horiz_sse2, vp9_convolve8_avg_horiz_sse2,
vp9_convolve8_vert_sse2, vp9_convolve8_avg_vert_sse2,
- vp9_convolve8_sse2, vp9_convolve8_avg_sse2);
+ vp9_convolve8_sse2, vp9_convolve8_avg_sse2, 0);
INSTANTIATE_TEST_CASE_P(SSE2, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_sse2),
@@ -626,14 +1536,14 @@
make_tuple(64, 32, &convolve8_sse2),
make_tuple(32, 64, &convolve8_sse2),
make_tuple(64, 64, &convolve8_sse2)));
-#endif
+#endif // CONFIG_VP9_HIGHBITDEPTH
#endif
#if HAVE_SSSE3
const ConvolveFunctions convolve8_ssse3(
vp9_convolve8_horiz_ssse3, vp9_convolve8_avg_horiz_ssse3,
vp9_convolve8_vert_ssse3, vp9_convolve8_avg_vert_ssse3,
- vp9_convolve8_ssse3, vp9_convolve8_avg_ssse3);
+ vp9_convolve8_ssse3, vp9_convolve8_avg_ssse3, 0);
INSTANTIATE_TEST_CASE_P(SSSE3, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_ssse3),
@@ -655,7 +1565,7 @@
const ConvolveFunctions convolve8_avx2(
vp9_convolve8_horiz_avx2, vp9_convolve8_avg_horiz_ssse3,
vp9_convolve8_vert_avx2, vp9_convolve8_avg_vert_ssse3,
- vp9_convolve8_avx2, vp9_convolve8_avg_ssse3);
+ vp9_convolve8_avx2, vp9_convolve8_avg_ssse3, 0);
INSTANTIATE_TEST_CASE_P(AVX2, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_avx2),
@@ -677,7 +1587,7 @@
const ConvolveFunctions convolve8_neon(
vp9_convolve8_horiz_neon, vp9_convolve8_avg_horiz_neon,
vp9_convolve8_vert_neon, vp9_convolve8_avg_vert_neon,
- vp9_convolve8_neon, vp9_convolve8_avg_neon);
+ vp9_convolve8_neon, vp9_convolve8_avg_neon, 0);
INSTANTIATE_TEST_CASE_P(NEON, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_neon),
@@ -699,7 +1609,7 @@
const ConvolveFunctions convolve8_dspr2(
vp9_convolve8_horiz_dspr2, vp9_convolve8_avg_horiz_dspr2,
vp9_convolve8_vert_dspr2, vp9_convolve8_avg_vert_dspr2,
- vp9_convolve8_dspr2, vp9_convolve8_avg_dspr2);
+ vp9_convolve8_dspr2, vp9_convolve8_avg_dspr2, 0);
INSTANTIATE_TEST_CASE_P(DSPR2, ConvolveTest, ::testing::Values(
make_tuple(4, 4, &convolve8_dspr2),
diff --git a/vp9/common/vp9_convolve.c b/vp9/common/vp9_convolve.c
index d8aaf32..ad70e59 100644
--- a/vp9/common/vp9_convolve.c
+++ b/vp9/common/vp9_convolve.c
@@ -282,3 +282,280 @@
dst += dst_stride;
}
}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const InterpKernel *x_filters,
+ int x0_q4, int x_step_q4,
+ int w, int h, int bd) {
+ int x, y;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ src -= SUBPEL_TAPS / 2 - 1;
+ for (y = 0; y < h; ++y) {
+ int x_q4 = x0_q4;
+ for (x = 0; x < w; ++x) {
+ const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+ const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+ int k, sum = 0;
+ for (k = 0; k < SUBPEL_TAPS; ++k)
+ sum += src_x[k] * x_filter[k];
+ dst[x] = clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+ x_q4 += x_step_q4;
+ }
+ src += src_stride;
+ dst += dst_stride;
+ }
+}
+
+static void high_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const InterpKernel *x_filters,
+ int x0_q4, int x_step_q4,
+ int w, int h, int bd) {
+ int x, y;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ src -= SUBPEL_TAPS / 2 - 1;
+ for (y = 0; y < h; ++y) {
+ int x_q4 = x0_q4;
+ for (x = 0; x < w; ++x) {
+ const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+ const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+ int k, sum = 0;
+ for (k = 0; k < SUBPEL_TAPS; ++k)
+ sum += src_x[k] * x_filter[k];
+ dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+ clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+ x_q4 += x_step_q4;
+ }
+ src += src_stride;
+ dst += dst_stride;
+ }
+}
+
+static void high_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const InterpKernel *y_filters,
+ int y0_q4, int y_step_q4, int w, int h,
+ int bd) {
+ int x, y;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+ for (x = 0; x < w; ++x) {
+ int y_q4 = y0_q4;
+ for (y = 0; y < h; ++y) {
+ const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+ const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+ int k, sum = 0;
+ for (k = 0; k < SUBPEL_TAPS; ++k)
+ sum += src_y[k * src_stride] * y_filter[k];
+ dst[y * dst_stride] = clip_pixel_high(
+ ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+ y_q4 += y_step_q4;
+ }
+ ++src;
+ ++dst;
+ }
+}
+
+static void high_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const InterpKernel *y_filters,
+ int y0_q4, int y_step_q4, int w, int h,
+ int bd) {
+ int x, y;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+ for (x = 0; x < w; ++x) {
+ int y_q4 = y0_q4;
+ for (y = 0; y < h; ++y) {
+ const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+ const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+ int k, sum = 0;
+ for (k = 0; k < SUBPEL_TAPS; ++k)
+ sum += src_y[k * src_stride] * y_filter[k];
+ dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+ clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+ y_q4 += y_step_q4;
+ }
+ ++src;
+ ++dst;
+ }
+}
+
+static void high_convolve(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const InterpKernel *const x_filters,
+ int x0_q4, int x_step_q4,
+ const InterpKernel *const y_filters,
+ int y0_q4, int y_step_q4,
+ int w, int h, int bd) {
+ // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+ // 2d filtering proceeds in 2 steps:
+ // (1) Interpolate horizontally into an intermediate buffer, temp.
+ // (2) Interpolate temp vertically to derive the sub-pixel result.
+ // Deriving the maximum number of rows in the temp buffer (135):
+ // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+ // --Largest block size is 64x64 pixels.
+ // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+ // original frame (in 1/16th pixel units).
+ // --Must round-up because block may be located at sub-pixel position.
+ // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+ // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+ uint16_t temp[64 * 135];
+ int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
+
+ assert(w <= 64);
+ assert(h <= 64);
+ assert(y_step_q4 <= 32);
+ assert(x_step_q4 <= 32);
+
+ if (intermediate_height < h)
+ intermediate_height = h;
+
+ high_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+ src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+ x_filters, x0_q4, x_step_q4, w,
+ intermediate_height, bd);
+ high_convolve_vert(CONVERT_TO_BYTEPTR(temp) + 64 * (SUBPEL_TAPS / 2 - 1),
+ 64, dst, dst_stride, y_filters, y0_q4, y_step_q4,
+ w, h, bd);
+}
+
+
+void vp9_high_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+ (void)filter_y;
+ (void)y_step_q4;
+
+ high_convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+ x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+ (void)filter_y;
+ (void)y_step_q4;
+
+ high_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+ x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ const InterpKernel *const filters_y = get_filter_base(filter_y);
+ const int y0_q4 = get_filter_offset(filter_y, filters_y);
+ (void)filter_x;
+ (void)x_step_q4;
+
+ high_convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+ y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ const InterpKernel *const filters_y = get_filter_base(filter_y);
+ const int y0_q4 = get_filter_offset(filter_y, filters_y);
+ (void)filter_x;
+ (void)x_step_q4;
+
+ high_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+ y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ const InterpKernel *const filters_x = get_filter_base(filter_x);
+ const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+ const InterpKernel *const filters_y = get_filter_base(filter_y);
+ const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+ high_convolve(src, src_stride, dst, dst_stride,
+ filters_x, x0_q4, x_step_q4,
+ filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd) {
+ // Fixed size intermediate buffer places limits on parameters.
+ DECLARE_ALIGNED_ARRAY(16, uint16_t, temp, 64 * 64);
+ assert(w <= 64);
+ assert(h <= 64);
+
+ vp9_high_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+ filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+ vp9_high_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), 64, dst, dst_stride,
+ NULL, 0, NULL, 0, w, h, bd);
+}
+
+void vp9_high_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int filter_x_stride,
+ const int16_t *filter_y, int filter_y_stride,
+ int w, int h, int bd) {
+ int r;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ (void)filter_x;
+ (void)filter_y;
+ (void)filter_x_stride;
+ (void)filter_y_stride;
+ (void)bd;
+
+ for (r = h; r > 0; --r) {
+ vpx_memcpy(dst, src, w * sizeof(uint16_t));
+ src += src_stride;
+ dst += dst_stride;
+ }
+}
+
+void vp9_high_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+ uint8_t *dst8, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int filter_x_stride,
+ const int16_t *filter_y, int filter_y_stride,
+ int w, int h, int bd) {
+ int x, y;
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+ (void)filter_x;
+ (void)filter_y;
+ (void)filter_x_stride;
+ (void)filter_y_stride;
+ (void)bd;
+
+ for (y = 0; y < h; ++y) {
+ for (x = 0; x < w; ++x) {
+ dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+ }
+ src += src_stride;
+ dst += dst_stride;
+ }
+}
+#endif
diff --git a/vp9/common/vp9_convolve.h b/vp9/common/vp9_convolve.h
index 6bf71fc..faf70b1 100644
--- a/vp9/common/vp9_convolve.h
+++ b/vp9/common/vp9_convolve.h
@@ -23,6 +23,14 @@
const int16_t *filter_y, int y_step_q4,
int w, int h);
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*high_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+ uint8_t *dst, ptrdiff_t dst_stride,
+ const int16_t *filter_x, int x_step_q4,
+ const int16_t *filter_y, int y_step_q4,
+ int w, int h, int bd);
+#endif
+
#ifdef __cplusplus
} // extern "C"
#endif
diff --git a/vp9/common/vp9_reconinter.c b/vp9/common/vp9_reconinter.c
index 86ae648..28723f6 100644
--- a/vp9/common/vp9_reconinter.c
+++ b/vp9/common/vp9_reconinter.c
@@ -63,6 +63,53 @@
} while (--b_h);
}
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_build_mc_border(const uint8_t *src8, int src_stride,
+ uint16_t *dst, int dst_stride,
+ int x, int y, int b_w, int b_h,
+ int w, int h) {
+ // Get a pointer to the start of the real data for this row.
+ const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+ const uint16_t *ref_row = src - x - y * src_stride;
+
+ if (y >= h)
+ ref_row += (h - 1) * src_stride;
+ else if (y > 0)
+ ref_row += y * src_stride;
+
+ do {
+ int right = 0, copy;
+ int left = x < 0 ? -x : 0;
+
+ if (left > b_w)
+ left = b_w;
+
+ if (x + b_w > w)
+ right = x + b_w - w;
+
+ if (right > b_w)
+ right = b_w;
+
+ copy = b_w - left - right;
+
+ if (left)
+ vpx_memset16(dst, ref_row[0], left);
+
+ if (copy)
+ memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+
+ if (right)
+ vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+
+ dst += dst_stride;
+ ++y;
+
+ if (y > 0 && y < h)
+ ref_row += src_stride;
+ } while (--b_h);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
static void inter_predictor(const uint8_t *src, int src_stride,
uint8_t *dst, int dst_stride,
const int subpel_x,
@@ -97,6 +144,42 @@
sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4);
}
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_inter_predictor(const uint8_t *src, int src_stride,
+ uint8_t *dst, int dst_stride,
+ const int subpel_x,
+ const int subpel_y,
+ const struct scale_factors *sf,
+ int w, int h, int ref,
+ const InterpKernel *kernel,
+ int xs, int ys, int bd) {
+ sf->high_predict[subpel_x != 0][subpel_y != 0][ref](
+ src, src_stride, dst, dst_stride,
+ kernel[subpel_x], xs, kernel[subpel_y], ys, w, h, bd);
+}
+
+void vp9_high_build_inter_predictor(const uint8_t *src, int src_stride,
+ uint8_t *dst, int dst_stride,
+ const MV *src_mv,
+ const struct scale_factors *sf,
+ int w, int h, int ref,
+ const InterpKernel *kernel,
+ enum mv_precision precision,
+ int x, int y, int bd) {
+ const int is_q4 = precision == MV_PRECISION_Q4;
+ const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+ is_q4 ? src_mv->col : src_mv->col * 2 };
+ MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+ const int subpel_x = mv.col & SUBPEL_MASK;
+ const int subpel_y = mv.row & SUBPEL_MASK;
+
+ src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
+
+ high_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+ sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4, bd);
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
static INLINE int round_mv_comp_q4(int value) {
return (value < 0 ? value - 2 : value + 2) / 4;
}
@@ -222,8 +305,19 @@
pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride
+ (scaled_mv.col >> SUBPEL_BITS);
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+ high_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+ subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys,
+ xd->bd);
+ } else {
+ inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+ subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+ }
+#else
inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif // CONFIG_VP9_HIGHBITDEPTH
}
}
@@ -393,16 +487,64 @@
y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
// Extend the border.
- build_mc_border(buf_ptr1, pre_buf->stride, xd->mc_buf, x1 - x0 + 1,
- x0, y0, x1 - x0 + 1, y1 - y0 + 1, frame_width,
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+ high_build_mc_border(buf_ptr1,
+ pre_buf->stride,
+ xd->mc_buf_high,
+ x1 - x0 + 1,
+ x0,
+ y0,
+ x1 - x0 + 1,
+ y1 - y0 + 1,
+ frame_width,
+ frame_height);
+ buf_stride = x1 - x0 + 1;
+ buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
+ y_pad * 3 * buf_stride + x_pad * 3;
+ } else {
+ build_mc_border(buf_ptr1,
+ pre_buf->stride,
+ xd->mc_buf,
+ x1 - x0 + 1,
+ x0,
+ y0,
+ x1 - x0 + 1,
+ y1 - y0 + 1,
+ frame_width,
+ frame_height);
+ buf_stride = x1 - x0 + 1;
+ buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+ }
+#else
+ build_mc_border(buf_ptr1,
+ pre_buf->stride,
+ xd->mc_buf,
+ x1 - x0 + 1,
+ x0,
+ y0,
+ x1 - x0 + 1,
+ y1 - y0 + 1,
+ frame_width,
frame_height);
buf_stride = x1 - x0 + 1;
buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+#endif // CONFIG_VP9_HIGHBITDEPTH
}
}
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+ high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+ subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+ } else {
+ inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+ subpel_y, sf, w, h, ref, kernel, xs, ys);
+ }
+#else
inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif // CONFIG_VP9_HIGHBITDEPTH
}
}
diff --git a/vp9/common/vp9_reconinter.h b/vp9/common/vp9_reconinter.h
index 58c596e..e70cc4c 100644
--- a/vp9/common/vp9_reconinter.h
+++ b/vp9/common/vp9_reconinter.h
@@ -39,6 +39,17 @@
enum mv_precision precision,
int x, int y);
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_high_build_inter_predictor(const uint8_t *src, int src_stride,
+ uint8_t *dst, int dst_stride,
+ const MV *mv_q3,
+ const struct scale_factors *sf,
+ int w, int h, int do_avg,
+ const InterpKernel *kernel,
+ enum mv_precision precision,
+ int x, int y, int bd);
+#endif
+
static INLINE int scaled_buffer_offset(int x_offset, int y_offset, int stride,
const struct scale_factors *sf) {
const int x = sf ? sf->scale_value_x(x_offset, sf) : x_offset;
diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index b75ea64..0f52ae1 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -607,6 +607,33 @@
specialize qw/vp9_high_dc_128_predictor_32x32/;
#
+ # Sub Pixel Filters
+ #
+ add_proto qw/void vp9_high_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve_copy/;
+
+ add_proto qw/void vp9_high_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve_avg/;
+
+ add_proto qw/void vp9_high_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8/, "$sse2_x86_64";
+
+ add_proto qw/void vp9_high_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8_horiz/, "$sse2_x86_64";
+
+ add_proto qw/void vp9_high_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8_vert/, "$sse2_x86_64";
+
+ add_proto qw/void vp9_high_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8_avg/, "$sse2_x86_64";
+
+ add_proto qw/void vp9_high_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8_avg_horiz/, "$sse2_x86_64";
+
+ add_proto qw/void vp9_high_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+ specialize qw/vp9_high_convolve8_avg_vert/, "$sse2_x86_64";
+
+ #
# dct
#
add_proto qw/void vp9_high_idct4x4_1_add/, "const tran_low_t *input, uint8_t *dest, int dest_stride, int bd";
diff --git a/vp9/common/vp9_scale.c b/vp9/common/vp9_scale.c
index 2f58323..63e2b53 100644
--- a/vp9/common/vp9_scale.c
+++ b/vp9/common/vp9_scale.c
@@ -43,9 +43,16 @@
return res;
}
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+ int other_w, int other_h,
+ int this_w, int this_h,
+ int use_high) {
+#else
void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
int other_w, int other_h,
int this_w, int this_h) {
+#endif
if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
sf->x_scale_fp = REF_INVALID_SCALE;
sf->y_scale_fp = REF_INVALID_SCALE;
@@ -111,4 +118,48 @@
// 2D subpel motion always gets filtered in both directions
sf->predict[1][1][0] = vp9_convolve8;
sf->predict[1][1][1] = vp9_convolve8_avg;
+#if CONFIG_VP9_HIGHBITDEPTH
+ if (use_high) {
+ if (sf->x_step_q4 == 16) {
+ if (sf->y_step_q4 == 16) {
+ // No scaling in either direction.
+ sf->high_predict[0][0][0] = vp9_high_convolve_copy;
+ sf->high_predict[0][0][1] = vp9_high_convolve_avg;
+ sf->high_predict[0][1][0] = vp9_high_convolve8_vert;
+ sf->high_predict[0][1][1] = vp9_high_convolve8_avg_vert;
+ sf->high_predict[1][0][0] = vp9_high_convolve8_horiz;
+ sf->high_predict[1][0][1] = vp9_high_convolve8_avg_horiz;
+ } else {
+ // No scaling in x direction. Must always scale in the y direction.
+ sf->high_predict[0][0][0] = vp9_high_convolve8_vert;
+ sf->high_predict[0][0][1] = vp9_high_convolve8_avg_vert;
+ sf->high_predict[0][1][0] = vp9_high_convolve8_vert;
+ sf->high_predict[0][1][1] = vp9_high_convolve8_avg_vert;
+ sf->high_predict[1][0][0] = vp9_high_convolve8;
+ sf->high_predict[1][0][1] = vp9_high_convolve8_avg;
+ }
+ } else {
+ if (sf->y_step_q4 == 16) {
+ // No scaling in the y direction. Must always scale in the x direction.
+ sf->high_predict[0][0][0] = vp9_high_convolve8_horiz;
+ sf->high_predict[0][0][1] = vp9_high_convolve8_avg_horiz;
+ sf->high_predict[0][1][0] = vp9_high_convolve8;
+ sf->high_predict[0][1][1] = vp9_high_convolve8_avg;
+ sf->high_predict[1][0][0] = vp9_high_convolve8_horiz;
+ sf->high_predict[1][0][1] = vp9_high_convolve8_avg_horiz;
+ } else {
+ // Must always scale in both directions.
+ sf->high_predict[0][0][0] = vp9_high_convolve8;
+ sf->high_predict[0][0][1] = vp9_high_convolve8_avg;
+ sf->high_predict[0][1][0] = vp9_high_convolve8;
+ sf->high_predict[0][1][1] = vp9_high_convolve8_avg;
+ sf->high_predict[1][0][0] = vp9_high_convolve8;
+ sf->high_predict[1][0][1] = vp9_high_convolve8_avg;
+ }
+ }
+ // 2D subpel motion always gets filtered in both directions.
+ sf->high_predict[1][1][0] = vp9_high_convolve8;
+ sf->high_predict[1][1][1] = vp9_high_convolve8_avg;
+ }
+#endif
}
diff --git a/vp9/common/vp9_scale.h b/vp9/common/vp9_scale.h
index ad6f5d7..2e923db 100644
--- a/vp9/common/vp9_scale.h
+++ b/vp9/common/vp9_scale.h
@@ -32,13 +32,23 @@
int (*scale_value_y)(int val, const struct scale_factors *sf);
convolve_fn_t predict[2][2][2]; // horiz, vert, avg
+#if CONFIG_VP9_HIGHBITDEPTH
+ high_convolve_fn_t high_predict[2][2][2]; // horiz, vert, avg
+#endif
};
MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf);
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+ int other_w, int other_h,
+ int this_w, int this_h,
+ int use_high);
+#else
void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
int other_w, int other_h,
int this_w, int this_h);
+#endif
static INLINE int vp9_is_valid_scale(const struct scale_factors *sf) {
return sf->x_scale_fp != REF_INVALID_SCALE &&
diff --git a/vp9/common/x86/vp9_asm_stubs.c b/vp9/common/x86/vp9_asm_stubs.c
index b6847b9..407573a 100644
--- a/vp9/common/x86/vp9_asm_stubs.c
+++ b/vp9/common/x86/vp9_asm_stubs.c
@@ -139,6 +139,153 @@
filter_x, x_step_q4, filter_y, y_step_q4, w, h); \
} \
}
+
+#if CONFIG_VP9_HIGHBITDEPTH
+
+typedef void high_filter8_1dfunction (
+ const uint16_t *src_ptr,
+ const ptrdiff_t src_pitch,
+ uint16_t *output_ptr,
+ ptrdiff_t out_pitch,
+ unsigned int output_height,
+ const int16_t *filter,
+ int bd
+);
+
+#define HIGH_FUN_CONV_1D(name, step_q4, filter, dir, src_start, avg, opt) \
+ void vp9_high_convolve8_##name##_##opt(const uint8_t *src8, \
+ ptrdiff_t src_stride, \
+ uint8_t *dst8, ptrdiff_t dst_stride, \
+ const int16_t *filter_x, \
+ int x_step_q4, \
+ const int16_t *filter_y, \
+ int y_step_q4, \
+ int w, int h, int bd) { \
+ if (step_q4 == 16 && filter[3] != 128) { \
+ uint16_t *src = CONVERT_TO_SHORTPTR(src8); \
+ uint16_t *dst = CONVERT_TO_SHORTPTR(dst8); \
+ if (filter[0] || filter[1] || filter[2]) { \
+ while (w >= 16) { \
+ vp9_high_filter_block1d16_##dir##8_##avg##opt(src_start, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 16; \
+ dst += 16; \
+ w -= 16; \
+ } \
+ while (w >= 8) { \
+ vp9_high_filter_block1d8_##dir##8_##avg##opt(src_start, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 8; \
+ dst += 8; \
+ w -= 8; \
+ } \
+ while (w >= 4) { \
+ vp9_high_filter_block1d4_##dir##8_##avg##opt(src_start, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 4; \
+ dst += 4; \
+ w -= 4; \
+ } \
+ } else { \
+ while (w >= 16) { \
+ vp9_high_filter_block1d16_##dir##2_##avg##opt(src, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 16; \
+ dst += 16; \
+ w -= 16; \
+ } \
+ while (w >= 8) { \
+ vp9_high_filter_block1d8_##dir##2_##avg##opt(src, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 8; \
+ dst += 8; \
+ w -= 8; \
+ } \
+ while (w >= 4) { \
+ vp9_high_filter_block1d4_##dir##2_##avg##opt(src, \
+ src_stride, \
+ dst, \
+ dst_stride, \
+ h, \
+ filter, \
+ bd); \
+ src += 4; \
+ dst += 4; \
+ w -= 4; \
+ } \
+ } \
+ } \
+ if (w) { \
+ vp9_high_convolve8_##name##_c(src8, src_stride, dst8, dst_stride, \
+ filter_x, x_step_q4, filter_y, y_step_q4, \
+ w, h, bd); \
+ } \
+}
+
+#define HIGH_FUN_CONV_2D(avg, opt) \
+void vp9_high_convolve8_##avg##opt(const uint8_t *src, ptrdiff_t src_stride, \
+ uint8_t *dst, ptrdiff_t dst_stride, \
+ const int16_t *filter_x, int x_step_q4, \
+ const int16_t *filter_y, int y_step_q4, \
+ int w, int h, int bd) { \
+ assert(w <= 64); \
+ assert(h <= 64); \
+ if (x_step_q4 == 16 && y_step_q4 == 16) { \
+ if (filter_x[0] || filter_x[1] || filter_x[2] || filter_x[3] == 128 || \
+ filter_y[0] || filter_y[1] || filter_y[2] || filter_y[3] == 128) { \
+ DECLARE_ALIGNED_ARRAY(16, uint16_t, fdata2, 64 * 71); \
+ vp9_high_convolve8_horiz_##opt(src - 3 * src_stride, src_stride, \
+ CONVERT_TO_BYTEPTR(fdata2), 64, \
+ filter_x, x_step_q4, filter_y, y_step_q4, \
+ w, h + 7, bd); \
+ vp9_high_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2) + 192, \
+ 64, dst, dst_stride, \
+ filter_x, x_step_q4, filter_y, \
+ y_step_q4, w, h, bd); \
+ } else { \
+ DECLARE_ALIGNED_ARRAY(16, uint16_t, fdata2, 64 * 65); \
+ vp9_high_convolve8_horiz_##opt(src, src_stride, \
+ CONVERT_TO_BYTEPTR(fdata2), 64, \
+ filter_x, x_step_q4, filter_y, y_step_q4, \
+ w, h + 1, bd); \
+ vp9_high_convolve8_##avg##vert_##opt(CONVERT_TO_BYTEPTR(fdata2), 64, \
+ dst, dst_stride, \
+ filter_x, x_step_q4, filter_y, \
+ y_step_q4, w, h, bd); \
+ } \
+ } else { \
+ vp9_high_convolve8_##avg##c(src, src_stride, dst, dst_stride, \
+ filter_x, x_step_q4, filter_y, y_step_q4, w, \
+ h, bd); \
+ } \
+}
+#endif // CONFIG_VP9_HIGHBITDEPTH
+
#if HAVE_AVX2 && HAVE_SSSE3
filter8_1dfunction vp9_filter_block1d16_v8_avx2;
filter8_1dfunction vp9_filter_block1d16_h8_avx2;
@@ -336,4 +483,75 @@
// int w, int h);
FUN_CONV_2D(, sse2);
FUN_CONV_2D(avg_ , sse2);
+
+#if CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
+high_filter8_1dfunction vp9_high_filter_block1d16_v8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_h8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_v8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_h8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_v8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_h8_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_v8_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_h8_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_v8_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_h8_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_v8_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_h8_avg_sse2;
+
+high_filter8_1dfunction vp9_high_filter_block1d16_v2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_h2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_v2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_h2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_v2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_h2_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_v2_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d16_h2_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_v2_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d8_h2_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_v2_avg_sse2;
+high_filter8_1dfunction vp9_high_filter_block1d4_h2_avg_sse2;
+
+// void vp9_high_convolve8_horiz_sse2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h, int bd);
+// void vp9_high_convolve8_vert_sse2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h, int bd);
+// void vp9_high_convolve8_avg_horiz_sse2(const uint8_t *src,
+// ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x,
+// int x_step_q4,
+// const int16_t *filter_y,
+// int y_step_q4,
+// int w, int h, int bd);
+// void vp9_high_convolve8_avg_vert_sse2(const uint8_t *src,
+// ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h, int bd);
+HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , sse2);
+HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , sse2);
+HIGH_FUN_CONV_1D(avg_horiz, x_step_q4, filter_x, h, src, avg_, sse2);
+HIGH_FUN_CONV_1D(avg_vert, y_step_q4, filter_y, v, src - src_stride * 3, avg_,
+ sse2);
+
+// void vp9_high_convolve8_sse2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h, int bd);
+// void vp9_high_convolve8_avg_sse2(const uint8_t *src, ptrdiff_t src_stride,
+// uint8_t *dst, ptrdiff_t dst_stride,
+// const int16_t *filter_x, int x_step_q4,
+// const int16_t *filter_y, int y_step_q4,
+// int w, int h, int bd);
+HIGH_FUN_CONV_2D(, sse2);
+HIGH_FUN_CONV_2D(avg_ , sse2);
+#endif // CONFIG_VP9_HIGHBITDEPTH && ARCH_X86_64
#endif // HAVE_SSE2
diff --git a/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm b/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm
new file mode 100644
index 0000000..4bdbb83
--- /dev/null
+++ b/vp9/common/x86/vp9_high_subpixel_8t_sse2.asm
@@ -0,0 +1,962 @@
+;
+; Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+
+%include "vpx_ports/x86_abi_support.asm"
+
+;Note: tap3 and tap4 have to be applied and added after other taps to avoid
+;overflow.
+
+%macro HIGH_GET_FILTERS_4 0
+ mov rdx, arg(5) ;filter ptr
+ mov rcx, 0x00000040
+
+ movdqa xmm7, [rdx] ;load filters
+ pshuflw xmm0, xmm7, 0b ;k0
+ pshuflw xmm1, xmm7, 01010101b ;k1
+ pshuflw xmm2, xmm7, 10101010b ;k2
+ pshuflw xmm3, xmm7, 11111111b ;k3
+ psrldq xmm7, 8
+ pshuflw xmm4, xmm7, 0b ;k4
+ pshuflw xmm5, xmm7, 01010101b ;k5
+ pshuflw xmm6, xmm7, 10101010b ;k6
+ pshuflw xmm7, xmm7, 11111111b ;k7
+
+ punpcklwd xmm0, xmm6
+ punpcklwd xmm2, xmm5
+ punpcklwd xmm3, xmm4
+ punpcklwd xmm1, xmm7
+
+ movdqa k0k6, xmm0
+ movdqa k2k5, xmm2
+ movdqa k3k4, xmm3
+ movdqa k1k7, xmm1
+
+ movq xmm6, rcx
+ pshufd xmm6, xmm6, 0
+ movdqa krd, xmm6
+
+ ;Compute max and min values of a pixel
+ mov rdx, 0x00010001
+ movsxd rcx, DWORD PTR arg(6) ;bps
+ movq xmm0, rdx
+ movq xmm1, rcx
+ pshufd xmm0, xmm0, 0b
+ movdqa xmm2, xmm0
+ psllw xmm0, xmm1
+ psubw xmm0, xmm2
+ pxor xmm1, xmm1
+ movdqa max, xmm0 ;max value (for clamping)
+ movdqa min, xmm1 ;min value (for clamping)
+
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+ punpcklwd xmm0, xmm6 ;two row in one register
+ punpcklwd xmm1, xmm7
+ punpcklwd xmm2, xmm5
+ punpcklwd xmm3, xmm4
+
+ pmaddwd xmm0, k0k6 ;multiply the filter factors
+ pmaddwd xmm1, k1k7
+ pmaddwd xmm2, k2k5
+ pmaddwd xmm3, k3k4
+
+ paddd xmm0, xmm1 ;sum
+ paddd xmm0, xmm2
+ paddd xmm0, xmm3
+
+ paddd xmm0, krd ;rounding
+ psrad xmm0, 7 ;shift
+ packssdw xmm0, xmm0 ;pack to word
+
+ ;clamp the values
+ pminsw xmm0, max
+ pmaxsw xmm0, min
+
+%if %1
+ movq xmm1, [rdi]
+ pavgw xmm0, xmm1
+%endif
+ movq [rdi], xmm0
+%endm
+
+%macro HIGH_GET_FILTERS 0
+ mov rdx, arg(5) ;filter ptr
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+ mov rcx, 0x00000040
+
+ movdqa xmm7, [rdx] ;load filters
+ pshuflw xmm0, xmm7, 0b ;k0
+ pshuflw xmm1, xmm7, 01010101b ;k1
+ pshuflw xmm2, xmm7, 10101010b ;k2
+ pshuflw xmm3, xmm7, 11111111b ;k3
+ pshufhw xmm4, xmm7, 0b ;k4
+ pshufhw xmm5, xmm7, 01010101b ;k5
+ pshufhw xmm6, xmm7, 10101010b ;k6
+ pshufhw xmm7, xmm7, 11111111b ;k7
+ punpcklqdq xmm2, xmm2
+ punpcklqdq xmm3, xmm3
+ punpcklwd xmm0, xmm1
+ punpckhwd xmm6, xmm7
+ punpckhwd xmm2, xmm5
+ punpckhwd xmm3, xmm4
+
+ movdqa k0k1, xmm0 ;store filter factors on stack
+ movdqa k6k7, xmm6
+ movdqa k2k5, xmm2
+ movdqa k3k4, xmm3
+
+ movq xmm6, rcx
+ pshufd xmm6, xmm6, 0
+ movdqa krd, xmm6 ;rounding
+
+ ;Compute max and min values of a pixel
+ mov rdx, 0x00010001
+ movsxd rcx, DWORD PTR arg(6) ;bps
+ movq xmm0, rdx
+ movq xmm1, rcx
+ pshufd xmm0, xmm0, 0b
+ movdqa xmm2, xmm0
+ psllw xmm0, xmm1
+ psubw xmm0, xmm2
+ pxor xmm1, xmm1
+ movdqa max, xmm0 ;max value (for clamping)
+ movdqa min, xmm1 ;min value (for clamping)
+%endm
+
+%macro LOAD_VERT_8 1
+ movdqu xmm0, [rsi + %1] ;0
+ movdqu xmm1, [rsi + rax + %1] ;1
+ movdqu xmm6, [rsi + rdx * 2 + %1] ;6
+ lea rsi, [rsi + rax]
+ movdqu xmm7, [rsi + rdx * 2 + %1] ;7
+ movdqu xmm2, [rsi + rax + %1] ;2
+ movdqu xmm3, [rsi + rax * 2 + %1] ;3
+ movdqu xmm4, [rsi + rdx + %1] ;4
+ movdqu xmm5, [rsi + rax * 4 + %1] ;5
+%endm
+
+%macro HIGH_APPLY_FILTER_8 2
+ movdqu temp, xmm4
+ movdqa xmm4, xmm0
+ punpcklwd xmm0, xmm1
+ punpckhwd xmm4, xmm1
+ movdqa xmm1, xmm6
+ punpcklwd xmm6, xmm7
+ punpckhwd xmm1, xmm7
+ movdqa xmm7, xmm2
+ punpcklwd xmm2, xmm5
+ punpckhwd xmm7, xmm5
+
+ movdqu xmm5, temp
+ movdqu temp, xmm4
+ movdqa xmm4, xmm3
+ punpcklwd xmm3, xmm5
+ punpckhwd xmm4, xmm5
+ movdqu xmm5, temp
+
+ pmaddwd xmm0, k0k1
+ pmaddwd xmm5, k0k1
+ pmaddwd xmm6, k6k7
+ pmaddwd xmm1, k6k7
+ pmaddwd xmm2, k2k5
+ pmaddwd xmm7, k2k5
+ pmaddwd xmm3, k3k4
+ pmaddwd xmm4, k3k4
+
+ paddd xmm0, xmm6
+ paddd xmm0, xmm2
+ paddd xmm0, xmm3
+ paddd xmm5, xmm1
+ paddd xmm5, xmm7
+ paddd xmm5, xmm4
+
+ paddd xmm0, krd ;rounding
+ paddd xmm5, krd
+ psrad xmm0, 7 ;shift
+ psrad xmm5, 7
+ packssdw xmm0, xmm5 ;pack back to word
+
+ ;clamp the values
+ pminsw xmm0, max
+ pmaxsw xmm0, min
+
+%if %1
+ movdqu xmm1, [rdi + %2]
+ pavgw xmm0, xmm1
+%endif
+ movdqu [rdi + %2], xmm0
+%endm
+
+;void vp9_filter_block1d4_v8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pitch,
+; unsigned char *output_ptr,
+; unsigned int out_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d4_v8_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_v8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 7
+ %define k0k6 [rsp + 16 * 0]
+ %define k2k5 [rsp + 16 * 1]
+ %define k3k4 [rsp + 16 * 2]
+ %define k1k7 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define max [rsp + 16 * 5]
+ %define min [rsp + 16 * 6]
+
+ HIGH_GET_FILTERS_4
+
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movq xmm0, [rsi] ;load src: row 0
+ movq xmm1, [rsi + rax] ;1
+ movq xmm6, [rsi + rdx * 2] ;6
+ lea rsi, [rsi + rax]
+ movq xmm7, [rsi + rdx * 2] ;7
+ movq xmm2, [rsi + rax] ;2
+ movq xmm3, [rsi + rax * 2] ;3
+ movq xmm4, [rsi + rdx] ;4
+ movq xmm5, [rsi + rax * 4] ;5
+
+ HIGH_APPLY_FILTER_4 0
+
+ lea rdi, [rdi + rbx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 7
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+;void vp9_filter_block1d8_v8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pitch,
+; unsigned char *output_ptr,
+; unsigned int out_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d8_v8_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_v8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ LOAD_VERT_8 0
+ HIGH_APPLY_FILTER_8 0, 0
+
+ lea rdi, [rdi + rbx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+;void vp9_filter_block1d16_v8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pitch,
+; unsigned char *output_ptr,
+; unsigned int out_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d16_v8_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_v8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ LOAD_VERT_8 0
+ HIGH_APPLY_FILTER_8 0, 0
+ sub rsi, rax
+
+ LOAD_VERT_8 16
+ HIGH_APPLY_FILTER_8 0, 16
+ add rdi, rbx
+
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d4_v8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_v8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 7
+ %define k0k6 [rsp + 16 * 0]
+ %define k2k5 [rsp + 16 * 1]
+ %define k3k4 [rsp + 16 * 2]
+ %define k1k7 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define max [rsp + 16 * 5]
+ %define min [rsp + 16 * 6]
+
+ HIGH_GET_FILTERS_4
+
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movq xmm0, [rsi] ;load src: row 0
+ movq xmm1, [rsi + rax] ;1
+ movq xmm6, [rsi + rdx * 2] ;6
+ lea rsi, [rsi + rax]
+ movq xmm7, [rsi + rdx * 2] ;7
+ movq xmm2, [rsi + rax] ;2
+ movq xmm3, [rsi + rax * 2] ;3
+ movq xmm4, [rsi + rdx] ;4
+ movq xmm5, [rsi + rax * 4] ;5
+
+ HIGH_APPLY_FILTER_4 1
+
+ lea rdi, [rdi + rbx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 7
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d8_v8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_v8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+.loop:
+ LOAD_VERT_8 0
+ HIGH_APPLY_FILTER_8 1, 0
+
+ lea rdi, [rdi + rbx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_v8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_v8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ push rbx
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rbx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rbx, [rbx + rbx]
+ lea rdx, [rax + rax * 2]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+.loop:
+ LOAD_VERT_8 0
+ HIGH_APPLY_FILTER_8 1, 0
+ sub rsi, rax
+
+ LOAD_VERT_8 16
+ HIGH_APPLY_FILTER_8 1, 16
+ add rdi, rbx
+
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+ pop rbx
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+;void vp9_filter_block1d4_h8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pixels_per_line,
+; unsigned char *output_ptr,
+; unsigned int output_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d4_h8_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_h8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 7
+ %define k0k6 [rsp + 16 * 0]
+ %define k2k5 [rsp + 16 * 1]
+ %define k3k4 [rsp + 16 * 2]
+ %define k1k7 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define max [rsp + 16 * 5]
+ %define min [rsp + 16 * 6]
+
+ HIGH_GET_FILTERS_4
+
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm4, [rsi + 2]
+ movdqa xmm1, xmm0
+ movdqa xmm6, xmm4
+ movdqa xmm7, xmm4
+ movdqa xmm2, xmm0
+ movdqa xmm3, xmm0
+ movdqa xmm5, xmm4
+
+ psrldq xmm1, 2
+ psrldq xmm6, 4
+ psrldq xmm7, 6
+ psrldq xmm2, 4
+ psrldq xmm3, 6
+ psrldq xmm5, 2
+
+ HIGH_APPLY_FILTER_4 0
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 7
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+;void vp9_filter_block1d8_h8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pixels_per_line,
+; unsigned char *output_ptr,
+; unsigned int output_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d8_h8_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_h8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm1, [rsi - 4]
+ movdqu xmm2, [rsi - 2]
+ movdqu xmm3, [rsi]
+ movdqu xmm4, [rsi + 2]
+ movdqu xmm5, [rsi + 4]
+ movdqu xmm6, [rsi + 6]
+ movdqu xmm7, [rsi + 8]
+
+ HIGH_APPLY_FILTER_8 0, 0
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+;void vp9_filter_block1d16_h8_sse2
+;(
+; unsigned char *src_ptr,
+; unsigned int src_pixels_per_line,
+; unsigned char *output_ptr,
+; unsigned int output_pitch,
+; unsigned int output_height,
+; short *filter
+;)
+global sym(vp9_high_filter_block1d16_h8_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_h8_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm1, [rsi - 4]
+ movdqu xmm2, [rsi - 2]
+ movdqu xmm3, [rsi]
+ movdqu xmm4, [rsi + 2]
+ movdqu xmm5, [rsi + 4]
+ movdqu xmm6, [rsi + 6]
+ movdqu xmm7, [rsi + 8]
+
+ HIGH_APPLY_FILTER_8 0, 0
+
+ movdqu xmm0, [rsi + 10] ;load src
+ movdqu xmm1, [rsi + 12]
+ movdqu xmm2, [rsi + 14]
+ movdqu xmm3, [rsi + 16]
+ movdqu xmm4, [rsi + 18]
+ movdqu xmm5, [rsi + 20]
+ movdqu xmm6, [rsi + 22]
+ movdqu xmm7, [rsi + 24]
+
+ HIGH_APPLY_FILTER_8 0, 16
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d4_h8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_h8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 7
+ %define k0k6 [rsp + 16 * 0]
+ %define k2k5 [rsp + 16 * 1]
+ %define k3k4 [rsp + 16 * 2]
+ %define k1k7 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define max [rsp + 16 * 5]
+ %define min [rsp + 16 * 6]
+
+ HIGH_GET_FILTERS_4
+
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm4, [rsi + 2]
+ movdqa xmm1, xmm0
+ movdqa xmm6, xmm4
+ movdqa xmm7, xmm4
+ movdqa xmm2, xmm0
+ movdqa xmm3, xmm0
+ movdqa xmm5, xmm4
+
+ psrldq xmm1, 2
+ psrldq xmm6, 4
+ psrldq xmm7, 6
+ psrldq xmm2, 4
+ psrldq xmm3, 6
+ psrldq xmm5, 2
+
+ HIGH_APPLY_FILTER_4 1
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 7
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d8_h8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_h8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm1, [rsi - 4]
+ movdqu xmm2, [rsi - 2]
+ movdqu xmm3, [rsi]
+ movdqu xmm4, [rsi + 2]
+ movdqu xmm5, [rsi + 4]
+ movdqu xmm6, [rsi + 6]
+ movdqu xmm7, [rsi + 8]
+
+ HIGH_APPLY_FILTER_8 1, 0
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_h8_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_h8_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ ALIGN_STACK 16, rax
+ sub rsp, 16 * 8
+ %define k0k1 [rsp + 16 * 0]
+ %define k6k7 [rsp + 16 * 1]
+ %define k2k5 [rsp + 16 * 2]
+ %define k3k4 [rsp + 16 * 3]
+ %define krd [rsp + 16 * 4]
+ %define temp [rsp + 16 * 5]
+ %define max [rsp + 16 * 6]
+ %define min [rsp + 16 * 7]
+
+ HIGH_GET_FILTERS
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ lea rax, [rax + rax] ;bytes per line
+ lea rdx, [rdx + rdx]
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+
+.loop:
+ movdqu xmm0, [rsi - 6] ;load src
+ movdqu xmm1, [rsi - 4]
+ movdqu xmm2, [rsi - 2]
+ movdqu xmm3, [rsi]
+ movdqu xmm4, [rsi + 2]
+ movdqu xmm5, [rsi + 4]
+ movdqu xmm6, [rsi + 6]
+ movdqu xmm7, [rsi + 8]
+
+ HIGH_APPLY_FILTER_8 1, 0
+
+ movdqu xmm0, [rsi + 10] ;load src
+ movdqu xmm1, [rsi + 12]
+ movdqu xmm2, [rsi + 14]
+ movdqu xmm3, [rsi + 16]
+ movdqu xmm4, [rsi + 18]
+ movdqu xmm5, [rsi + 20]
+ movdqu xmm6, [rsi + 22]
+ movdqu xmm7, [rsi + 24]
+
+ HIGH_APPLY_FILTER_8 1, 16
+
+ lea rsi, [rsi + rax]
+ lea rdi, [rdi + rdx]
+ dec rcx
+ jnz .loop
+
+ add rsp, 16 * 8
+ pop rsp
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
diff --git a/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm b/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm
new file mode 100644
index 0000000..b7d4a61
--- /dev/null
+++ b/vp9/common/x86/vp9_high_subpixel_bilinear_sse2.asm
@@ -0,0 +1,494 @@
+;
+; Copyright (c) 2014 The WebM project authors. All Rights Reserved.
+;
+; Use of this source code is governed by a BSD-style license
+; that can be found in the LICENSE file in the root of the source
+; tree. An additional intellectual property rights grant can be found
+; in the file PATENTS. All contributing project authors may
+; be found in the AUTHORS file in the root of the source tree.
+;
+
+%include "vpx_ports/x86_abi_support.asm"
+
+%macro HIGH_GET_PARAM_4 0
+ mov rdx, arg(5) ;filter ptr
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+ mov rcx, 0x00000040
+
+ movdqa xmm3, [rdx] ;load filters
+ pshuflw xmm4, xmm3, 11111111b ;k3
+ psrldq xmm3, 8
+ pshuflw xmm3, xmm3, 0b ;k4
+ punpcklwd xmm4, xmm3 ;k3k4
+
+ movq xmm3, rcx ;rounding
+ pshufd xmm3, xmm3, 0
+
+ mov rdx, 0x00010001
+ movsxd rcx, DWORD PTR arg(6) ;bps
+ movq xmm5, rdx
+ movq xmm2, rcx
+ pshufd xmm5, xmm5, 0b
+ movdqa xmm1, xmm5
+ psllw xmm5, xmm2
+ psubw xmm5, xmm1 ;max value (for clamping)
+ pxor xmm2, xmm2 ;min value (for clamping)
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_4 1
+
+ punpcklwd xmm0, xmm1 ;two row in one register
+ pmaddwd xmm0, xmm4 ;multiply the filter factors
+
+ paddd xmm0, xmm3 ;rounding
+ psrad xmm0, 7 ;shift
+ packssdw xmm0, xmm0 ;pack to word
+
+ ;clamp the values
+ pminsw xmm0, xmm5
+ pmaxsw xmm0, xmm2
+
+%if %1
+ movq xmm1, [rdi]
+ pavgw xmm0, xmm1
+%endif
+
+ movq [rdi], xmm0
+ lea rsi, [rsi + 2*rax]
+ lea rdi, [rdi + 2*rdx]
+ dec rcx
+%endm
+
+%if ARCH_X86_64
+%macro HIGH_GET_PARAM 0
+ mov rdx, arg(5) ;filter ptr
+ mov rsi, arg(0) ;src_ptr
+ mov rdi, arg(2) ;output_ptr
+ mov rcx, 0x00000040
+
+ movdqa xmm6, [rdx] ;load filters
+
+ pshuflw xmm7, xmm6, 11111111b ;k3
+ pshufhw xmm6, xmm6, 0b ;k4
+ psrldq xmm6, 8
+ punpcklwd xmm7, xmm6 ;k3k4k3k4k3k4k3k4
+
+ movq xmm4, rcx ;rounding
+ pshufd xmm4, xmm4, 0
+
+ mov rdx, 0x00010001
+ movsxd rcx, DWORD PTR arg(6) ;bps
+ movq xmm8, rdx
+ movq xmm5, rcx
+ pshufd xmm8, xmm8, 0b
+ movdqa xmm1, xmm8
+ psllw xmm8, xmm5
+ psubw xmm8, xmm1 ;max value (for clamping)
+ pxor xmm5, xmm5 ;min value (for clamping)
+
+ movsxd rax, DWORD PTR arg(1) ;pixels_per_line
+ movsxd rdx, DWORD PTR arg(3) ;out_pitch
+ movsxd rcx, DWORD PTR arg(4) ;output_height
+%endm
+
+%macro HIGH_APPLY_FILTER_8 1
+ movdqa xmm6, xmm0
+ punpckhwd xmm6, xmm1
+ punpcklwd xmm0, xmm1
+ pmaddwd xmm6, xmm7
+ pmaddwd xmm0, xmm7
+
+ paddd xmm6, xmm4 ;rounding
+ paddd xmm0, xmm4 ;rounding
+ psrad xmm6, 7 ;shift
+ psrad xmm0, 7 ;shift
+ packssdw xmm0, xmm6 ;pack back to word
+
+ ;clamp the values
+ pminsw xmm0, xmm8
+ pmaxsw xmm0, xmm5
+
+%if %1
+ movdqu xmm1, [rdi]
+ pavgw xmm0, xmm1
+%endif
+ movdqu [rdi], xmm0 ;store the result
+
+ lea rsi, [rsi + 2*rax]
+ lea rdi, [rdi + 2*rdx]
+ dec rcx
+%endm
+
+%macro HIGH_APPLY_FILTER_16 1
+ movdqa xmm9, xmm0
+ movdqa xmm6, xmm2
+ punpckhwd xmm9, xmm1
+ punpckhwd xmm6, xmm3
+ punpcklwd xmm0, xmm1
+ punpcklwd xmm2, xmm3
+
+ pmaddwd xmm9, xmm7
+ pmaddwd xmm6, xmm7
+ pmaddwd xmm0, xmm7
+ pmaddwd xmm2, xmm7
+
+ paddd xmm9, xmm4 ;rounding
+ paddd xmm6, xmm4
+ paddd xmm0, xmm4
+ paddd xmm2, xmm4
+
+ psrad xmm9, 7 ;shift
+ psrad xmm6, 7
+ psrad xmm0, 7
+ psrad xmm2, 7
+
+ packssdw xmm0, xmm9 ;pack back to word
+ packssdw xmm2, xmm6 ;pack back to word
+
+ ;clamp the values
+ pminsw xmm0, xmm8
+ pmaxsw xmm0, xmm5
+ pminsw xmm2, xmm8
+ pmaxsw xmm2, xmm5
+
+%if %1
+ movdqu xmm1, [rdi]
+ movdqu xmm3, [rdi + 16]
+ pavgw xmm0, xmm1
+ pavgw xmm2, xmm3
+%endif
+ movdqu [rdi], xmm0 ;store the result
+ movdqu [rdi + 16], xmm2 ;store the result
+
+ lea rsi, [rsi + 2*rax]
+ lea rdi, [rdi + 2*rdx]
+ dec rcx
+%endm
+%endif
+
+global sym(vp9_high_filter_block1d4_v2_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_v2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM_4
+.loop:
+ movq xmm0, [rsi] ;load src
+ movq xmm1, [rsi + 2*rax]
+
+ HIGH_APPLY_FILTER_4 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+%if ARCH_X86_64
+global sym(vp9_high_filter_block1d8_v2_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_v2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 8
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;0
+ movdqu xmm1, [rsi + 2*rax] ;1
+
+ HIGH_APPLY_FILTER_8 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_v2_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_v2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 9
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;0
+ movdqu xmm2, [rsi + 16]
+ movdqu xmm1, [rsi + 2*rax] ;1
+ movdqu xmm3, [rsi + 2*rax + 16]
+
+ HIGH_APPLY_FILTER_16 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+%endif
+
+global sym(vp9_high_filter_block1d4_v2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_v2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM_4
+.loop:
+ movq xmm0, [rsi] ;load src
+ movq xmm1, [rsi + 2*rax]
+
+ HIGH_APPLY_FILTER_4 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+%if ARCH_X86_64
+global sym(vp9_high_filter_block1d8_v2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_v2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 8
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;0
+ movdqu xmm1, [rsi + 2*rax] ;1
+
+ HIGH_APPLY_FILTER_8 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_v2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_v2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 9
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;0
+ movdqu xmm1, [rsi + 2*rax] ;1
+ movdqu xmm2, [rsi + 16]
+ movdqu xmm3, [rsi + 2*rax + 16]
+
+ HIGH_APPLY_FILTER_16 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+%endif
+
+global sym(vp9_high_filter_block1d4_h2_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_h2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM_4
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqa xmm1, xmm0
+ psrldq xmm1, 2
+
+ HIGH_APPLY_FILTER_4 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+%if ARCH_X86_64
+global sym(vp9_high_filter_block1d8_h2_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_h2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 8
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqu xmm1, [rsi + 2]
+
+ HIGH_APPLY_FILTER_8 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_h2_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_h2_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 9
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqu xmm1, [rsi + 2]
+ movdqu xmm2, [rsi + 16]
+ movdqu xmm3, [rsi + 18]
+
+ HIGH_APPLY_FILTER_16 0
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+%endif
+
+global sym(vp9_high_filter_block1d4_h2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d4_h2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM_4
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqa xmm1, xmm0
+ psrldq xmm1, 2
+
+ HIGH_APPLY_FILTER_4 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+%if ARCH_X86_64
+global sym(vp9_high_filter_block1d8_h2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d8_h2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 8
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqu xmm1, [rsi + 2]
+
+ HIGH_APPLY_FILTER_8 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+
+global sym(vp9_high_filter_block1d16_h2_avg_sse2) PRIVATE
+sym(vp9_high_filter_block1d16_h2_avg_sse2):
+ push rbp
+ mov rbp, rsp
+ SHADOW_ARGS_TO_STACK 7
+ SAVE_XMM 9
+ push rsi
+ push rdi
+ ; end prolog
+
+ HIGH_GET_PARAM
+.loop:
+ movdqu xmm0, [rsi] ;load src
+ movdqu xmm1, [rsi + 2]
+ movdqu xmm2, [rsi + 16]
+ movdqu xmm3, [rsi + 18]
+
+ HIGH_APPLY_FILTER_16 1
+ jnz .loop
+
+ ; begin epilog
+ pop rdi
+ pop rsi
+ RESTORE_XMM
+ UNSHADOW_ARGS
+ pop rbp
+ ret
+%endif
diff --git a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c b/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
index 448ad5a..de2394b 100644
--- a/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
+++ b/vp9/common/x86/vp9_loopfilter_intrin_sse2.c
@@ -12,6 +12,10 @@
#include "vp9/common/vp9_loopfilter.h"
#include "vpx_ports/emmintrin_compat.h"
+static INLINE __m128i abs_diff(__m128i a, __m128i b) {
+ return _mm_or_si128(_mm_subs_epu8(a, b), _mm_subs_epu8(b, a));
+}
+
static void mb_lpf_horizontal_edge_w_sse2_8(unsigned char *s,
int p,
const unsigned char *_blimit,
@@ -46,15 +50,12 @@
{
__m128i abs_p1q1, abs_p0q0, abs_q1q0, fe, ff, work;
- abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
- _mm_subs_epu8(q0p0, q1p1));
+ abs_p1p0 = abs_diff(q1p1, q0p0);
abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
fe = _mm_set1_epi8(0xfe);
ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
- _mm_subs_epu8(p0q0, q0p0));
- abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
- _mm_subs_epu8(p1q1, q1p1));
+ abs_p0q0 = abs_diff(q0p0, p0q0);
+ abs_p1q1 = abs_diff(q1p1, p1q1);
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
@@ -68,10 +69,8 @@
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
- _mm_subs_epu8(q1p1, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
- _mm_subs_epu8(q2p2, q3p3)));
+ work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+ abs_diff(q3p3, q2p2));
mask = _mm_max_epu8(work, mask);
mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
mask = _mm_subs_epu8(mask, limit);
@@ -125,10 +124,7 @@
{
__m128i work;
- flat = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
- _mm_subs_epu8(q0p0, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
- _mm_subs_epu8(q0p0, q3p3)));
+ flat = _mm_max_epu8(abs_diff(q2p2, q0p0), abs_diff(q3p3, q0p0));
flat = _mm_max_epu8(abs_p1p0, flat);
flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
flat = _mm_subs_epu8(flat, one);
@@ -142,21 +138,12 @@
q6p6 = _mm_loadl_epi64((__m128i *)(s - 7 * p));
q6p6 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q6p6),
(__m64 *)(s + 6 * p)));
-
- flat2 = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q4p4, q0p0),
- _mm_subs_epu8(q0p0, q4p4)),
- _mm_or_si128(_mm_subs_epu8(q5p5, q0p0),
- _mm_subs_epu8(q0p0, q5p5)));
+ flat2 = _mm_max_epu8(abs_diff(q4p4, q0p0), abs_diff(q5p5, q0p0));
q7p7 = _mm_loadl_epi64((__m128i *)(s - 8 * p));
q7p7 = _mm_castps_si128(_mm_loadh_pi(_mm_castsi128_ps(q7p7),
(__m64 *)(s + 7 * p)));
-
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q6p6, q0p0),
- _mm_subs_epu8(q0p0, q6p6)),
- _mm_or_si128(_mm_subs_epu8(q7p7, q0p0),
- _mm_subs_epu8(q0p0, q7p7)));
-
+ work = _mm_max_epu8(abs_diff(q6p6, q0p0), abs_diff(q7p7, q0p0));
flat2 = _mm_max_epu8(work, flat2);
flat2 = _mm_max_epu8(flat2, _mm_srli_si128(flat2, 8));
flat2 = _mm_subs_epu8(flat2, one);
@@ -364,20 +351,34 @@
}
}
+static INLINE __m128i filter_add2_sub2(__m128i total, __m128i a1, __m128i a2,
+ __m128i s1, __m128i s2) {
+ total = _mm_add_epi16(a1, total);
+ total = _mm_add_epi16(_mm_sub_epi16(total, _mm_add_epi16(s1, s2)), a2);
+ return total;
+}
+
+static INLINE __m128i filter8_mask(__m128i flat, __m128i other_filt,
+ __m128i f8_lo, __m128i f8_hi) {
+ const __m128i f8 = _mm_packus_epi16(_mm_srli_epi16(f8_lo, 3),
+ _mm_srli_epi16(f8_hi, 3));
+ const __m128i result = _mm_and_si128(flat, f8);
+ return _mm_or_si128(_mm_andnot_si128(flat, other_filt), result);
+}
+
+static INLINE __m128i filter16_mask(__m128i flat, __m128i other_filt,
+ __m128i f_lo, __m128i f_hi) {
+ const __m128i f = _mm_packus_epi16(_mm_srli_epi16(f_lo, 4),
+ _mm_srli_epi16(f_hi, 4));
+ const __m128i result = _mm_and_si128(flat, f);
+ return _mm_or_si128(_mm_andnot_si128(flat, other_filt), result);
+}
+
static void mb_lpf_horizontal_edge_w_sse2_16(unsigned char *s,
int p,
const unsigned char *_blimit,
const unsigned char *_limit,
const unsigned char *_thresh) {
- DECLARE_ALIGNED_ARRAY(16, unsigned char, flat2_op, 7 * 16);
- DECLARE_ALIGNED_ARRAY(16, unsigned char, flat2_oq, 7 * 16);
-
- DECLARE_ALIGNED_ARRAY(16, unsigned char, flat_op, 3 * 16);
- DECLARE_ALIGNED_ARRAY(16, unsigned char, flat_oq, 3 * 16);
-
- DECLARE_ALIGNED_ARRAY(16, unsigned char, ap, 8 * 16);
- DECLARE_ALIGNED_ARRAY(16, unsigned char, aq, 8 * 16);
-
const __m128i zero = _mm_set1_epi16(0);
const __m128i one = _mm_set1_epi8(1);
const __m128i blimit = _mm_load_si128((const __m128i *)_blimit);
@@ -387,8 +388,14 @@
__m128i p7, p6, p5;
__m128i p4, p3, p2, p1, p0, q0, q1, q2, q3, q4;
__m128i q5, q6, q7;
- int i = 0;
+ __m128i op2, op1, op0, oq0, oq1, oq2;
+
+ __m128i max_abs_p1p0q1q0;
+
+ p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
+ p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
+ p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
p4 = _mm_loadu_si128((__m128i *)(s - 5 * p));
p3 = _mm_loadu_si128((__m128i *)(s - 4 * p));
p2 = _mm_loadu_si128((__m128i *)(s - 3 * p));
@@ -399,58 +406,59 @@
q2 = _mm_loadu_si128((__m128i *)(s + 2 * p));
q3 = _mm_loadu_si128((__m128i *)(s + 3 * p));
q4 = _mm_loadu_si128((__m128i *)(s + 4 * p));
-
- _mm_store_si128((__m128i *)&ap[4 * 16], p4);
- _mm_store_si128((__m128i *)&ap[3 * 16], p3);
- _mm_store_si128((__m128i *)&ap[2 * 16], p2);
- _mm_store_si128((__m128i *)&ap[1 * 16], p1);
- _mm_store_si128((__m128i *)&ap[0 * 16], p0);
- _mm_store_si128((__m128i *)&aq[4 * 16], q4);
- _mm_store_si128((__m128i *)&aq[3 * 16], q3);
- _mm_store_si128((__m128i *)&aq[2 * 16], q2);
- _mm_store_si128((__m128i *)&aq[1 * 16], q1);
- _mm_store_si128((__m128i *)&aq[0 * 16], q0);
-
+ q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
+ q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
+ q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
{
- const __m128i abs_p1p0 = _mm_or_si128(_mm_subs_epu8(p1, p0),
- _mm_subs_epu8(p0, p1));
- const __m128i abs_q1q0 = _mm_or_si128(_mm_subs_epu8(q1, q0),
- _mm_subs_epu8(q0, q1));
+ const __m128i abs_p1p0 = abs_diff(p1, p0);
+ const __m128i abs_q1q0 = abs_diff(q1, q0);
const __m128i fe = _mm_set1_epi8(0xfe);
- const __m128i ff = _mm_cmpeq_epi8(abs_p1p0, abs_p1p0);
- __m128i abs_p0q0 = _mm_or_si128(_mm_subs_epu8(p0, q0),
- _mm_subs_epu8(q0, p0));
- __m128i abs_p1q1 = _mm_or_si128(_mm_subs_epu8(p1, q1),
- _mm_subs_epu8(q1, p1));
+ const __m128i ff = _mm_cmpeq_epi8(zero, zero);
+ __m128i abs_p0q0 = abs_diff(p0, q0);
+ __m128i abs_p1q1 = abs_diff(p1, q1);
__m128i work;
- flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
- hev = _mm_subs_epu8(flat, thresh);
- hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+ max_abs_p1p0q1q0 = _mm_max_epu8(abs_p1p0, abs_q1q0);
abs_p0q0 =_mm_adds_epu8(abs_p0q0, abs_p0q0);
abs_p1q1 = _mm_srli_epi16(_mm_and_si128(abs_p1q1, fe), 1);
mask = _mm_subs_epu8(_mm_adds_epu8(abs_p0q0, abs_p1q1), blimit);
mask = _mm_xor_si128(_mm_cmpeq_epi8(mask, zero), ff);
// mask |= (abs(p0 - q0) * 2 + abs(p1 - q1) / 2 > blimit) * -1;
- mask = _mm_max_epu8(flat, mask);
+ mask = _mm_max_epu8(max_abs_p1p0q1q0, mask);
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p1),
- _mm_subs_epu8(p1, p2)),
- _mm_or_si128(_mm_subs_epu8(p3, p2),
- _mm_subs_epu8(p2, p3)));
+ work = _mm_max_epu8(abs_diff(p2, p1), abs_diff(p3, p2));
mask = _mm_max_epu8(work, mask);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2, q1),
- _mm_subs_epu8(q1, q2)),
- _mm_or_si128(_mm_subs_epu8(q3, q2),
- _mm_subs_epu8(q2, q3)));
+ work = _mm_max_epu8(abs_diff(q2, q1), abs_diff(q3, q2));
mask = _mm_max_epu8(work, mask);
mask = _mm_subs_epu8(mask, limit);
mask = _mm_cmpeq_epi8(mask, zero);
}
- // lp filter
+ {
+ __m128i work;
+ work = _mm_max_epu8(abs_diff(p2, p0), abs_diff(q2, q0));
+ flat = _mm_max_epu8(work, max_abs_p1p0q1q0);
+ work = _mm_max_epu8(abs_diff(p3, p0), abs_diff(q3, q0));
+ flat = _mm_max_epu8(work, flat);
+ work = _mm_max_epu8(abs_diff(p4, p0), abs_diff(q4, q0));
+ flat = _mm_subs_epu8(flat, one);
+ flat = _mm_cmpeq_epi8(flat, zero);
+ flat = _mm_and_si128(flat, mask);
+ flat2 = _mm_max_epu8(abs_diff(p5, p0), abs_diff(q5, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ work = _mm_max_epu8(abs_diff(p6, p0), abs_diff(q6, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ work = _mm_max_epu8(abs_diff(p7, p0), abs_diff(q7, q0));
+ flat2 = _mm_max_epu8(work, flat2);
+ flat2 = _mm_subs_epu8(flat2, one);
+ flat2 = _mm_cmpeq_epi8(flat2, zero);
+ flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ }
+
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // filter4
{
const __m128i t4 = _mm_set1_epi8(4);
const __m128i t3 = _mm_set1_epi8(3);
@@ -459,23 +467,27 @@
const __m128i t1f = _mm_set1_epi8(0x1f);
const __m128i t1 = _mm_set1_epi8(0x1);
const __m128i t7f = _mm_set1_epi8(0x7f);
+ const __m128i ff = _mm_cmpeq_epi8(t4, t4);
- __m128i ps1 = _mm_xor_si128(p1, t80);
- __m128i ps0 = _mm_xor_si128(p0, t80);
- __m128i qs0 = _mm_xor_si128(q0, t80);
- __m128i qs1 = _mm_xor_si128(q1, t80);
__m128i filt;
__m128i work_a;
__m128i filter1, filter2;
- filt = _mm_and_si128(_mm_subs_epi8(ps1, qs1), hev);
- work_a = _mm_subs_epi8(qs0, ps0);
+ op1 = _mm_xor_si128(p1, t80);
+ op0 = _mm_xor_si128(p0, t80);
+ oq0 = _mm_xor_si128(q0, t80);
+ oq1 = _mm_xor_si128(q1, t80);
+
+ hev = _mm_subs_epu8(max_abs_p1p0q1q0, thresh);
+ hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
+ filt = _mm_and_si128(_mm_subs_epi8(op1, oq1), hev);
+
+ work_a = _mm_subs_epi8(oq0, op0);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
filt = _mm_adds_epi8(filt, work_a);
// (vp9_filter + 3 * (qs0 - ps0)) & mask
filt = _mm_and_si128(filt, mask);
-
filter1 = _mm_adds_epi8(filt, t4);
filter2 = _mm_adds_epi8(filt, t3);
@@ -485,7 +497,7 @@
work_a = _mm_and_si128(work_a, te0);
filter1 = _mm_and_si128(filter1, t1f);
filter1 = _mm_or_si128(filter1, work_a);
- qs0 = _mm_xor_si128(_mm_subs_epi8(qs0, filter1), t80);
+ oq0 = _mm_xor_si128(_mm_subs_epi8(oq0, filter1), t80);
// Filter2 >> 3
work_a = _mm_cmpgt_epi8(zero, filter2);
@@ -493,7 +505,7 @@
work_a = _mm_and_si128(work_a, te0);
filter2 = _mm_and_si128(filter2, t1f);
filter2 = _mm_or_si128(filter2, work_a);
- ps0 = _mm_xor_si128(_mm_adds_epi8(ps0, filter2), t80);
+ op0 = _mm_xor_si128(_mm_adds_epi8(op0, filter2), t80);
// filt >> 1
filt = _mm_adds_epi8(filter1, t1);
@@ -503,345 +515,195 @@
filt = _mm_and_si128(filt, t7f);
filt = _mm_or_si128(filt, work_a);
filt = _mm_andnot_si128(hev, filt);
- ps1 = _mm_xor_si128(_mm_adds_epi8(ps1, filt), t80);
- qs1 = _mm_xor_si128(_mm_subs_epi8(qs1, filt), t80);
+ op1 = _mm_xor_si128(_mm_adds_epi8(op1, filt), t80);
+ oq1 = _mm_xor_si128(_mm_subs_epi8(oq1, filt), t80);
// loopfilter done
+ // ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
+ // filter8
{
- __m128i work;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p2, p0),
- _mm_subs_epu8(p0, p2)),
- _mm_or_si128(_mm_subs_epu8(q2, q0),
- _mm_subs_epu8(q0, q2)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p3, p0),
- _mm_subs_epu8(p0, p3)),
- _mm_or_si128(_mm_subs_epu8(q3, q0),
- _mm_subs_epu8(q0, q3)));
- flat = _mm_max_epu8(work, flat);
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p4, p0),
- _mm_subs_epu8(p0, p4)),
- _mm_or_si128(_mm_subs_epu8(q4, q0),
- _mm_subs_epu8(q0, q4)));
- flat = _mm_subs_epu8(flat, one);
- flat = _mm_cmpeq_epi8(flat, zero);
- flat = _mm_and_si128(flat, mask);
+ const __m128i four = _mm_set1_epi16(4);
+ const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+ const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+ const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+ const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+ const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+ const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+ const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+ const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
- p5 = _mm_loadu_si128((__m128i *)(s - 6 * p));
- q5 = _mm_loadu_si128((__m128i *)(s + 5 * p));
- flat2 = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p5, p0),
- _mm_subs_epu8(p0, p5)),
- _mm_or_si128(_mm_subs_epu8(q5, q0),
- _mm_subs_epu8(q0, q5)));
- _mm_store_si128((__m128i *)&ap[5 * 16], p5);
- _mm_store_si128((__m128i *)&aq[5 * 16], q5);
- flat2 = _mm_max_epu8(work, flat2);
- p6 = _mm_loadu_si128((__m128i *)(s - 7 * p));
- q6 = _mm_loadu_si128((__m128i *)(s + 6 * p));
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p6, p0),
- _mm_subs_epu8(p0, p6)),
- _mm_or_si128(_mm_subs_epu8(q6, q0),
- _mm_subs_epu8(q0, q6)));
- _mm_store_si128((__m128i *)&ap[6 * 16], p6);
- _mm_store_si128((__m128i *)&aq[6 * 16], q6);
- flat2 = _mm_max_epu8(work, flat2);
+ const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+ const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+ const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+ const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+ const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+ const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+ const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+ const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+ __m128i f8_lo, f8_hi;
- p7 = _mm_loadu_si128((__m128i *)(s - 8 * p));
- q7 = _mm_loadu_si128((__m128i *)(s + 7 * p));
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(p7, p0),
- _mm_subs_epu8(p0, p7)),
- _mm_or_si128(_mm_subs_epu8(q7, q0),
- _mm_subs_epu8(q0, q7)));
- _mm_store_si128((__m128i *)&ap[7 * 16], p7);
- _mm_store_si128((__m128i *)&aq[7 * 16], q7);
- flat2 = _mm_max_epu8(work, flat2);
- flat2 = _mm_subs_epu8(flat2, one);
- flat2 = _mm_cmpeq_epi8(flat2, zero);
- flat2 = _mm_and_si128(flat2, flat); // flat2 & flat & mask
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, four),
+ _mm_add_epi16(p3_lo, p2_lo));
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f8_lo),
+ _mm_add_epi16(p2_lo, p1_lo));
+ f8_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f8_lo);
+
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, four),
+ _mm_add_epi16(p3_hi, p2_hi));
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f8_hi),
+ _mm_add_epi16(p2_hi, p1_hi));
+ f8_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f8_hi);
+
+ op2 = filter8_mask(flat, p2, f8_lo, f8_hi);
+
+ f8_lo = filter_add2_sub2(f8_lo, q1_lo, p1_lo, p2_lo, p3_lo);
+ f8_hi = filter_add2_sub2(f8_hi, q1_hi, p1_hi, p2_hi, p3_hi);
+ op1 = filter8_mask(flat, op1, f8_lo, f8_hi);
+
+ f8_lo = filter_add2_sub2(f8_lo, q2_lo, p0_lo, p1_lo, p3_lo);
+ f8_hi = filter_add2_sub2(f8_hi, q2_hi, p0_hi, p1_hi, p3_hi);
+ op0 = filter8_mask(flat, op0, f8_lo, f8_hi);
+
+ f8_lo = filter_add2_sub2(f8_lo, q3_lo, q0_lo, p0_lo, p3_lo);
+ f8_hi = filter_add2_sub2(f8_hi, q3_hi, q0_hi, p0_hi, p3_hi);
+ oq0 = filter8_mask(flat, oq0, f8_lo, f8_hi);
+
+ f8_lo = filter_add2_sub2(f8_lo, q3_lo, q1_lo, q0_lo, p2_lo);
+ f8_hi = filter_add2_sub2(f8_hi, q3_hi, q1_hi, q0_hi, p2_hi);
+ oq1 = filter8_mask(flat, oq1, f8_lo, f8_hi);
+
+ f8_lo = filter_add2_sub2(f8_lo, q3_lo, q2_lo, q1_lo, p1_lo);
+ f8_hi = filter_add2_sub2(f8_hi, q3_hi, q2_hi, q1_hi, p1_hi);
+ oq2 = filter8_mask(flat, q2, f8_lo, f8_hi);
}
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
- // flat and wide flat calculations
+ // wide flat calculations
{
const __m128i eight = _mm_set1_epi16(8);
- const __m128i four = _mm_set1_epi16(4);
- __m128i temp_flat2 = flat2;
- unsigned char *src = s;
- int i = 0;
- do {
- __m128i workp_shft;
- __m128i a, b, c;
+ const __m128i p7_lo = _mm_unpacklo_epi8(p7, zero);
+ const __m128i p6_lo = _mm_unpacklo_epi8(p6, zero);
+ const __m128i p5_lo = _mm_unpacklo_epi8(p5, zero);
+ const __m128i p4_lo = _mm_unpacklo_epi8(p4, zero);
+ const __m128i p3_lo = _mm_unpacklo_epi8(p3, zero);
+ const __m128i p2_lo = _mm_unpacklo_epi8(p2, zero);
+ const __m128i p1_lo = _mm_unpacklo_epi8(p1, zero);
+ const __m128i p0_lo = _mm_unpacklo_epi8(p0, zero);
+ const __m128i q0_lo = _mm_unpacklo_epi8(q0, zero);
+ const __m128i q1_lo = _mm_unpacklo_epi8(q1, zero);
+ const __m128i q2_lo = _mm_unpacklo_epi8(q2, zero);
+ const __m128i q3_lo = _mm_unpacklo_epi8(q3, zero);
+ const __m128i q4_lo = _mm_unpacklo_epi8(q4, zero);
+ const __m128i q5_lo = _mm_unpacklo_epi8(q5, zero);
+ const __m128i q6_lo = _mm_unpacklo_epi8(q6, zero);
+ const __m128i q7_lo = _mm_unpacklo_epi8(q7, zero);
- unsigned int off = i * 8;
- p7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[7 * 16] + off)),
- zero);
- p6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[6 * 16] + off)),
- zero);
- p5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[5 * 16] + off)),
- zero);
- p4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[4 * 16] + off)),
- zero);
- p3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[3 * 16] + off)),
- zero);
- p2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[2 * 16] + off)),
- zero);
- p1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[1 * 16] + off)),
- zero);
- p0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&ap[0 * 16] + off)),
- zero);
- q0 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[0 * 16] + off)),
- zero);
- q1 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[1 * 16] + off)),
- zero);
- q2 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[2 * 16] + off)),
- zero);
- q3 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[3 * 16] + off)),
- zero);
- q4 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[4 * 16] + off)),
- zero);
- q5 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[5 * 16] + off)),
- zero);
- q6 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[6 * 16] + off)),
- zero);
- q7 = _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(&aq[7 * 16] + off)),
- zero);
+ const __m128i p7_hi = _mm_unpackhi_epi8(p7, zero);
+ const __m128i p6_hi = _mm_unpackhi_epi8(p6, zero);
+ const __m128i p5_hi = _mm_unpackhi_epi8(p5, zero);
+ const __m128i p4_hi = _mm_unpackhi_epi8(p4, zero);
+ const __m128i p3_hi = _mm_unpackhi_epi8(p3, zero);
+ const __m128i p2_hi = _mm_unpackhi_epi8(p2, zero);
+ const __m128i p1_hi = _mm_unpackhi_epi8(p1, zero);
+ const __m128i p0_hi = _mm_unpackhi_epi8(p0, zero);
+ const __m128i q0_hi = _mm_unpackhi_epi8(q0, zero);
+ const __m128i q1_hi = _mm_unpackhi_epi8(q1, zero);
+ const __m128i q2_hi = _mm_unpackhi_epi8(q2, zero);
+ const __m128i q3_hi = _mm_unpackhi_epi8(q3, zero);
+ const __m128i q4_hi = _mm_unpackhi_epi8(q4, zero);
+ const __m128i q5_hi = _mm_unpackhi_epi8(q5, zero);
+ const __m128i q6_hi = _mm_unpackhi_epi8(q6, zero);
+ const __m128i q7_hi = _mm_unpackhi_epi8(q7, zero);
- c = _mm_sub_epi16(_mm_slli_epi16(p7, 3), p7); // p7 * 7
- c = _mm_add_epi16(_mm_slli_epi16(p6, 1), _mm_add_epi16(p4, c));
+ __m128i f_lo;
+ __m128i f_hi;
- b = _mm_add_epi16(_mm_add_epi16(p3, four), _mm_add_epi16(p3, p2));
- a = _mm_add_epi16(p3, _mm_add_epi16(p2, p1));
- a = _mm_add_epi16(_mm_add_epi16(p0, q0), a);
+ f_lo = _mm_sub_epi16(_mm_slli_epi16(p7_lo, 3), p7_lo); // p7 * 7
+ f_lo = _mm_add_epi16(_mm_slli_epi16(p6_lo, 1),
+ _mm_add_epi16(p4_lo, f_lo));
+ f_lo = _mm_add_epi16(_mm_add_epi16(p3_lo, f_lo),
+ _mm_add_epi16(p2_lo, p1_lo));
+ f_lo = _mm_add_epi16(_mm_add_epi16(p0_lo, q0_lo), f_lo);
+ f_lo = _mm_add_epi16(_mm_add_epi16(p5_lo, eight), f_lo);
- _mm_storel_epi64((__m128i *)&flat_op[2 * 16 + i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ f_hi = _mm_sub_epi16(_mm_slli_epi16(p7_hi, 3), p7_hi); // p7 * 7
+ f_hi = _mm_add_epi16(_mm_slli_epi16(p6_hi, 1),
+ _mm_add_epi16(p4_hi, f_hi));
+ f_hi = _mm_add_epi16(_mm_add_epi16(p3_hi, f_hi),
+ _mm_add_epi16(p2_hi, p1_hi));
+ f_hi = _mm_add_epi16(_mm_add_epi16(p0_hi, q0_hi), f_hi);
+ f_hi = _mm_add_epi16(_mm_add_epi16(p5_hi, eight), f_hi);
- c = _mm_add_epi16(_mm_add_epi16(p5, eight), c);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[6 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ p6 = filter16_mask(flat2, p6, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 7 * p), p6);
- a = _mm_add_epi16(q1, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p2)), p1);
- _mm_storel_epi64((__m128i *)&flat_op[1 * 16 + i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ f_lo = filter_add2_sub2(f_lo, q1_lo, p5_lo, p6_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q1_hi, p5_hi, p6_hi, p7_hi);
+ p5 = filter16_mask(flat2, p5, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 6 * p), p5);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p6)), p5);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[5 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q2_lo, p4_lo, p5_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q2_hi, p4_hi, p5_hi, p7_hi);
+ p4 = filter16_mask(flat2, p4, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 5 * p), p4);
- a = _mm_add_epi16(q2, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p1)), p0);
- _mm_storel_epi64((__m128i *)&flat_op[i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ f_lo = filter_add2_sub2(f_lo, q3_lo, p3_lo, p4_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q3_hi, p3_hi, p4_hi, p7_hi);
+ p3 = filter16_mask(flat2, p3, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 4 * p), p3);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p5)), p4);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[4 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q4_lo, p2_lo, p3_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q4_hi, p2_hi, p3_hi, p7_hi);
+ op2 = filter16_mask(flat2, op2, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 3 * p), op2);
- a = _mm_add_epi16(q3, a);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p3, p0)), q0);
- _mm_storel_epi64((__m128i *)&flat_oq[i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ f_lo = filter_add2_sub2(f_lo, q5_lo, p1_lo, p2_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q5_hi, p1_hi, p2_hi, p7_hi);
+ op1 = filter16_mask(flat2, op1, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 2 * p), op1);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p4)), p3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[3 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q6_lo, p0_lo, p1_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q6_hi, p0_hi, p1_hi, p7_hi);
+ op0 = filter16_mask(flat2, op0, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 1 * p), op0);
- b = _mm_add_epi16(q3, b);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p2, q0)), q1);
- _mm_storel_epi64((__m128i *)&flat_oq[16 + i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q0_lo, p0_lo, p7_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q0_hi, p0_hi, p7_hi);
+ oq0 = filter16_mask(flat2, oq0, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s - 0 * p), oq0);
- c = _mm_add_epi16(q4, c);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p3)), p2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[2 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q1_lo, p6_lo, q0_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q1_hi, p6_hi, q0_hi);
+ oq1 = filter16_mask(flat2, oq1, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 1 * p), oq1);
- b = _mm_add_epi16(q3, b);
- b = _mm_add_epi16(_mm_sub_epi16(b, _mm_add_epi16(p1, q1)), q2);
- _mm_storel_epi64((__m128i *)&flat_oq[2 * 16 + i * 8],
- _mm_packus_epi16(_mm_srli_epi16(_mm_add_epi16(a, b), 3)
- , b));
- a = _mm_add_epi16(q5, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p2)), p1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q2_lo, p5_lo, q1_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q2_hi, p5_hi, q1_hi);
+ oq2 = filter16_mask(flat2, oq2, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 2 * p), oq2);
- a = _mm_add_epi16(q6, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p1)), p0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_op[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q3_lo, p4_lo, q2_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q3_hi, p4_hi, q2_hi);
+ q3 = filter16_mask(flat2, q3, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 3 * p), q3);
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p7, p0)), q0);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q4_lo, p3_lo, q3_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q4_hi, p3_hi, q3_hi);
+ q4 = filter16_mask(flat2, q4, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 4 * p), q4);
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p6, q0)), q1);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q5_lo, p2_lo, q4_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q5_hi, p2_hi, q4_hi);
+ q5 = filter16_mask(flat2, q5, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 5 * p), q5);
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p5, q1)), q2);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[2 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p4, q2)), q3);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[3 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p3, q3)), q4);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[4 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p2, q4)), q5);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[5 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- a = _mm_add_epi16(q7, a);
- c = _mm_add_epi16(_mm_sub_epi16(c, _mm_add_epi16(p1, q5)), q6);
- workp_shft = _mm_srli_epi16(_mm_add_epi16(a, c), 4);
- _mm_storel_epi64((__m128i *)&flat2_oq[6 * 16 + i * 8],
- _mm_packus_epi16(workp_shft, workp_shft));
-
- temp_flat2 = _mm_srli_si128(temp_flat2, 8);
- src += 8;
- } while (++i < 2);
+ f_lo = filter_add2_sub2(f_lo, q7_lo, q6_lo, p1_lo, q5_lo);
+ f_hi = filter_add2_sub2(f_hi, q7_hi, q6_hi, p1_hi, q5_hi);
+ q6 = filter16_mask(flat2, q6, f_lo, f_hi);
+ _mm_storeu_si128((__m128i *)(s + 6 * p), q6);
}
// wide flat
// ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
-
- work_a = _mm_load_si128((__m128i *)&ap[2 * 16]);
- p2 = _mm_load_si128((__m128i *)&flat_op[2 * 16]);
- work_a = _mm_andnot_si128(flat, work_a);
- p2 = _mm_and_si128(flat, p2);
- p2 = _mm_or_si128(work_a, p2);
- _mm_store_si128((__m128i *)&flat_op[2 * 16], p2);
-
- p1 = _mm_load_si128((__m128i *)&flat_op[1 * 16]);
- work_a = _mm_andnot_si128(flat, ps1);
- p1 = _mm_and_si128(flat, p1);
- p1 = _mm_or_si128(work_a, p1);
- _mm_store_si128((__m128i *)&flat_op[1 * 16], p1);
-
- p0 = _mm_load_si128((__m128i *)&flat_op[0]);
- work_a = _mm_andnot_si128(flat, ps0);
- p0 = _mm_and_si128(flat, p0);
- p0 = _mm_or_si128(work_a, p0);
- _mm_store_si128((__m128i *)&flat_op[0], p0);
-
- q0 = _mm_load_si128((__m128i *)&flat_oq[0]);
- work_a = _mm_andnot_si128(flat, qs0);
- q0 = _mm_and_si128(flat, q0);
- q0 = _mm_or_si128(work_a, q0);
- _mm_store_si128((__m128i *)&flat_oq[0], q0);
-
- q1 = _mm_load_si128((__m128i *)&flat_oq[1 * 16]);
- work_a = _mm_andnot_si128(flat, qs1);
- q1 = _mm_and_si128(flat, q1);
- q1 = _mm_or_si128(work_a, q1);
- _mm_store_si128((__m128i *)&flat_oq[1 * 16], q1);
-
- work_a = _mm_load_si128((__m128i *)&aq[2 * 16]);
- q2 = _mm_load_si128((__m128i *)&flat_oq[2 * 16]);
- work_a = _mm_andnot_si128(flat, work_a);
- q2 = _mm_and_si128(flat, q2);
- q2 = _mm_or_si128(work_a, q2);
- _mm_store_si128((__m128i *)&flat_oq[2 * 16], q2);
-
- // write out op6 - op3
- {
- unsigned char *dst = (s - 7 * p);
- for (i = 6; i > 2; i--) {
- __m128i flat2_output;
- work_a = _mm_load_si128((__m128i *)&ap[i * 16]);
- flat2_output = _mm_load_si128((__m128i *)&flat2_op[i * 16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- flat2_output = _mm_and_si128(flat2, flat2_output);
- work_a = _mm_or_si128(work_a, flat2_output);
- _mm_storeu_si128((__m128i *)dst, work_a);
- dst += p;
- }
- }
-
- work_a = _mm_load_si128((__m128i *)&flat_op[2 * 16]);
- p2 = _mm_load_si128((__m128i *)&flat2_op[2 * 16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p2 = _mm_and_si128(flat2, p2);
- p2 = _mm_or_si128(work_a, p2);
- _mm_storeu_si128((__m128i *)(s - 3 * p), p2);
-
- work_a = _mm_load_si128((__m128i *)&flat_op[1 * 16]);
- p1 = _mm_load_si128((__m128i *)&flat2_op[1 * 16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p1 = _mm_and_si128(flat2, p1);
- p1 = _mm_or_si128(work_a, p1);
- _mm_storeu_si128((__m128i *)(s - 2 * p), p1);
-
- work_a = _mm_load_si128((__m128i *)&flat_op[0]);
- p0 = _mm_load_si128((__m128i *)&flat2_op[0]);
- work_a = _mm_andnot_si128(flat2, work_a);
- p0 = _mm_and_si128(flat2, p0);
- p0 = _mm_or_si128(work_a, p0);
- _mm_storeu_si128((__m128i *)(s - 1 * p), p0);
-
- work_a = _mm_load_si128((__m128i *)&flat_oq[0]);
- q0 = _mm_load_si128((__m128i *)&flat2_oq[0]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q0 = _mm_and_si128(flat2, q0);
- q0 = _mm_or_si128(work_a, q0);
- _mm_storeu_si128((__m128i *)(s - 0 * p), q0);
-
- work_a = _mm_load_si128((__m128i *)&flat_oq[1 * 16]);
- q1 = _mm_load_si128((__m128i *)&flat2_oq[16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q1 = _mm_and_si128(flat2, q1);
- q1 = _mm_or_si128(work_a, q1);
- _mm_storeu_si128((__m128i *)(s + 1 * p), q1);
-
- work_a = _mm_load_si128((__m128i *)&flat_oq[2 * 16]);
- q2 = _mm_load_si128((__m128i *)&flat2_oq[2 * 16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- q2 = _mm_and_si128(flat2, q2);
- q2 = _mm_or_si128(work_a, q2);
- _mm_storeu_si128((__m128i *)(s + 2 * p), q2);
-
- // write out oq3 - oq7
- {
- unsigned char *dst = (s + 3 * p);
- for (i = 3; i < 7; i++) {
- __m128i flat2_output;
- work_a = _mm_load_si128((__m128i *)&aq[i * 16]);
- flat2_output = _mm_load_si128((__m128i *)&flat2_oq[i * 16]);
- work_a = _mm_andnot_si128(flat2, work_a);
- flat2_output = _mm_and_si128(flat2, flat2_output);
- work_a = _mm_or_si128(work_a, flat2_output);
- _mm_storeu_si128((__m128i *)dst, work_a);
- dst += p;
- }
- }
}
}
@@ -893,14 +755,11 @@
const __m128i fe = _mm_set1_epi8(0xfe);
const __m128i ff = _mm_cmpeq_epi8(fe, fe);
__m128i abs_p1q1, abs_p0q0, abs_q1q0, abs_p1p0, work;
- abs_p1p0 = _mm_or_si128(_mm_subs_epu8(q1p1, q0p0),
- _mm_subs_epu8(q0p0, q1p1));
+ abs_p1p0 = abs_diff(q1p1, q0p0);
abs_q1q0 = _mm_srli_si128(abs_p1p0, 8);
- abs_p0q0 = _mm_or_si128(_mm_subs_epu8(q0p0, p0q0),
- _mm_subs_epu8(p0q0, q0p0));
- abs_p1q1 = _mm_or_si128(_mm_subs_epu8(q1p1, p1q1),
- _mm_subs_epu8(p1q1, q1p1));
+ abs_p0q0 = abs_diff(q0p0, p0q0);
+ abs_p1q1 = abs_diff(q1p1, p1q1);
flat = _mm_max_epu8(abs_p1p0, abs_q1q0);
hev = _mm_subs_epu8(flat, thresh);
hev = _mm_xor_si128(_mm_cmpeq_epi8(hev, zero), ff);
@@ -914,10 +773,8 @@
// mask |= (abs(p1 - p0) > limit) * -1;
// mask |= (abs(q1 - q0) > limit) * -1;
- work = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q1p1),
- _mm_subs_epu8(q1p1, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q2p2),
- _mm_subs_epu8(q2p2, q3p3)));
+ work = _mm_max_epu8(abs_diff(q2p2, q1p1),
+ abs_diff(q3p3, q2p2));
mask = _mm_max_epu8(work, mask);
mask = _mm_max_epu8(mask, _mm_srli_si128(mask, 8));
mask = _mm_subs_epu8(mask, limit);
@@ -925,10 +782,8 @@
// flat_mask4
- flat = _mm_max_epu8(_mm_or_si128(_mm_subs_epu8(q2p2, q0p0),
- _mm_subs_epu8(q0p0, q2p2)),
- _mm_or_si128(_mm_subs_epu8(q3p3, q0p0),
- _mm_subs_epu8(q0p0, q3p3)));
+ flat = _mm_max_epu8(abs_diff(q2p2, q0p0),
+ abs_diff(q3p3, q0p0));
flat = _mm_max_epu8(abs_p1p0, flat);
flat = _mm_max_epu8(flat, _mm_srli_si128(flat, 8));
flat = _mm_subs_epu8(flat, one);
diff --git a/vp9/decoder/vp9_decodeframe.c b/vp9/decoder/vp9_decodeframe.c
index f99fa7a..499fb4f 100644
--- a/vp9/decoder/vp9_decodeframe.c
+++ b/vp9/decoder/vp9_decodeframe.c
@@ -1265,10 +1265,18 @@
for (i = 0; i < REFS_PER_FRAME; ++i) {
RefBuffer *const ref_buf = &cm->frame_refs[i];
+#if CONFIG_VP9_HIGHBITDEPTH
+ vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+ ref_buf->buf->y_crop_width,
+ ref_buf->buf->y_crop_height,
+ cm->width, cm->height,
+ cm->use_highbitdepth);
+#else
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
ref_buf->buf->y_crop_width,
ref_buf->buf->y_crop_height,
cm->width, cm->height);
+#endif
if (vp9_is_scaled(&ref_buf->sf))
vp9_extend_frame_borders(ref_buf->buf);
}
diff --git a/vp9/encoder/vp9_encoder.c b/vp9/encoder/vp9_encoder.c
index 9545ba0..9a61b2a 100644
--- a/vp9/encoder/vp9_encoder.c
+++ b/vp9/encoder/vp9_encoder.c
@@ -2767,10 +2767,17 @@
RefBuffer *const ref_buf = &cm->frame_refs[ref_frame - 1];
ref_buf->buf = buf;
ref_buf->idx = idx;
+#if CONFIG_VP9_HIGHBITDEPTH
+ vp9_setup_scale_factors_for_frame(&ref_buf->sf,
+ buf->y_crop_width, buf->y_crop_height,
+ cm->width, cm->height,
+ (buf->flags & YV12_FLAG_HIGHBITDEPTH) ?
+ 1 : 0);
+#else
vp9_setup_scale_factors_for_frame(&ref_buf->sf,
buf->y_crop_width, buf->y_crop_height,
cm->width, cm->height);
-
+#endif
if (vp9_is_scaled(&ref_buf->sf))
vp9_extend_frame_borders(buf);
}
diff --git a/vp9/encoder/vp9_firstpass.c b/vp9/encoder/vp9_firstpass.c
index df82be5..71c98ec 100644
--- a/vp9/encoder/vp9_firstpass.c
+++ b/vp9/encoder/vp9_firstpass.c
@@ -35,26 +35,28 @@
#include "vp9/encoder/vp9_rd.h"
#include "vp9/encoder/vp9_variance.h"
-#define OUTPUT_FPF 0
+#define OUTPUT_FPF 0
+#define ARF_STATS_OUTPUT 0
-#define IIFACTOR 12.5
-#define IIKFACTOR1 12.5
-#define IIKFACTOR2 15.0
-#define RMAX 512.0
-#define GF_RMAX 96.0
-#define ERR_DIVISOR 150.0
-#define MIN_DECAY_FACTOR 0.1
-#define SVC_FACTOR_PT_LOW 0.45
-#define FACTOR_PT_LOW 0.5
-#define FACTOR_PT_HIGH 0.9
-
-#define KF_MB_INTRA_MIN 150
-#define GF_MB_INTRA_MIN 100
+#define BOOST_FACTOR 12.5
+#define ERR_DIVISOR 100.0
+#define FACTOR_PT_LOW 0.5
+#define FACTOR_PT_HIGH 0.9
+#define FIRST_PASS_Q 10.0
+#define GF_MAX_BOOST 96.0
+#define INTRA_MODE_PENALTY 1024
+#define KF_MAX_BOOST 128.0
+#define MIN_DECAY_FACTOR 0.01
+#define MIN_GF_INTERVAL 4
+#define MIN_KF_BOOST 300
+#define NEW_MV_MODE_PENALTY 32
+#define SVC_FACTOR_PT_LOW 0.45
#define DOUBLE_DIVIDE_CHECK(x) ((x) < 0 ? (x) - 0.000001 : (x) + 0.000001)
-#define MIN_KF_BOOST 300
-#define MIN_GF_INTERVAL 4
+#if ARF_STATS_OUTPUT
+unsigned int arf_count = 0;
+#endif
static void swap_yv12(YV12_BUFFER_CONFIG *a, YV12_BUFFER_CONFIG *b) {
YV12_BUFFER_CONFIG temp = *a;
@@ -299,7 +301,7 @@
int num00, tmp_err, n;
const BLOCK_SIZE bsize = xd->mi[0]->mbmi.sb_type;
vp9_variance_fn_ptr_t v_fn_ptr = cpi->fn_ptr[bsize];
- const int new_mv_mode_penalty = 256;
+ const int new_mv_mode_penalty = NEW_MV_MODE_PENALTY;
int step_param = 3;
int further_steps = (MAX_MVSEARCH_STEPS - 1) - step_param;
@@ -364,7 +366,7 @@
int i;
for (i = 0; i < QINDEX_RANGE; ++i)
- if (vp9_convert_qindex_to_q(i, bit_depth) >= 30.0)
+ if (vp9_convert_qindex_to_q(i, bit_depth) >= FIRST_PASS_Q)
break;
if (i == QINDEX_RANGE)
@@ -414,7 +416,7 @@
int mvcount = 0;
int intercount = 0;
int second_ref_count = 0;
- int intrapenalty = 256;
+ const int intrapenalty = INTRA_MODE_PENALTY;
int neutral_count = 0;
int new_mv_count = 0;
int sum_in_vectors = 0;
@@ -830,12 +832,18 @@
vp9_clear_system_state();
{
FIRSTPASS_STATS fps;
+ // The minimum error here insures some bit alocation to frames even
+ // in static regions. The allocation per MB declines for larger formats
+ // where the typical "real" energy per MB also falls.
+ // Initial estimate here uses sqrt(mbs) to define the min_err, where the
+ // number of mbs is propotional to image area.
+ const double min_err = 200 * sqrt(cm->MBs);
fps.frame = cm->current_video_frame;
fps.spatial_layer_id = cpi->svc.spatial_layer_id;
- fps.intra_error = (double)(intra_error >> 8);
- fps.coded_error = (double)(coded_error >> 8);
- fps.sr_coded_error = (double)(sr_coded_error >> 8);
+ fps.coded_error = (double)(coded_error >> 8) + min_err;
+ fps.sr_coded_error = (double)(sr_coded_error >> 8) + min_err;
+ fps.intra_error = (double)(intra_error >> 8) + min_err;
fps.count = 1.0;
fps.pcnt_inter = (double)intercount / cm->MBs;
fps.pcnt_second_ref = (double)second_ref_count / cm->MBs;
@@ -1033,17 +1041,6 @@
10000000.0);
}
- // Calculate a minimum intra value to be used in determining the IIratio
- // scores used in the second pass. We have this minimum to make sure
- // that clips that are static but "low complexity" in the intra domain
- // are still boosted appropriately for KF/GF/ARF.
- if (!is_two_pass_svc) {
- // We don't know the number of MBs for each layer at this point.
- // So we will do it later.
- twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
- twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
- }
-
// This variable monitors how far behind the second ref update is lagging.
twopass->sr_update_lag = 1;
@@ -1067,34 +1064,58 @@
// Reset the vbr bits off target counter
cpi->rc.vbr_bits_off_target = 0;
+
+ // Static sequence monitor variables.
+ twopass->kf_zeromotion_pct = 100;
+ twopass->last_kfgroup_zeromotion_pct = 100;
+}
+
+#define SR_DIFF_PART 0.0015
+#define MOTION_AMP_PART 0.003
+#define INTRA_PART 0.005
+#define DEFAULT_DECAY_LIMIT 0.75
+#define LOW_SR_DIFF_TRHESH 0.1
+#define SR_DIFF_MAX 128.0
+
+static double get_sr_decay_rate(const VP9_COMMON *cm,
+ const FIRSTPASS_STATS *frame) {
+ double sr_diff = (frame->sr_coded_error - frame->coded_error) / cm->MBs;
+ double sr_decay = 1.0;
+ const double motion_amplitude_factor =
+ frame->pcnt_motion * ((frame->mvc_abs + frame->mvr_abs) / 2);
+ const double pcnt_intra = 100 * (1.0 - frame->pcnt_inter);
+
+ if ((sr_diff > LOW_SR_DIFF_TRHESH)) {
+ sr_diff = MIN(sr_diff, SR_DIFF_MAX);
+ sr_decay = 1.0 - (SR_DIFF_PART * sr_diff) -
+ (MOTION_AMP_PART * motion_amplitude_factor) -
+ (INTRA_PART * pcnt_intra);
+ }
+ return MAX(sr_decay, MIN(DEFAULT_DECAY_LIMIT, frame->pcnt_inter));
}
// This function gives an estimate of how badly we believe the prediction
// quality is decaying from frame to frame.
-static double get_prediction_decay_rate(const VP9_COMMON *cm,
- const FIRSTPASS_STATS *next_frame) {
- // Look at the observed drop in prediction quality between the last frame
- // and the GF buffer (which contains an older frame).
- const double mb_sr_err_diff = (next_frame->sr_coded_error -
- next_frame->coded_error) / cm->MBs;
- const double second_ref_decay = mb_sr_err_diff <= 512.0
- ? fclamp(pow(1.0 - (mb_sr_err_diff / 512.0), 0.5), 0.85, 1.0)
- : 0.85;
-
- return MIN(second_ref_decay, next_frame->pcnt_inter);
-}
-
-// This function gives an estimate of how badly we believe the prediction
-// quality is decaying from frame to frame.
-static double get_zero_motion_factor(const FIRSTPASS_STATS *frame) {
- const double sr_ratio = frame->coded_error /
- DOUBLE_DIVIDE_CHECK(frame->sr_coded_error);
+static double get_zero_motion_factor(const VP9_COMMON *cm,
+ const FIRSTPASS_STATS *frame) {
const double zero_motion_pct = frame->pcnt_inter -
frame->pcnt_motion;
-
- return MIN(sr_ratio, zero_motion_pct);
+ double sr_decay = get_sr_decay_rate(cm, frame);
+ return MIN(sr_decay, zero_motion_pct);
}
+#define ZM_POWER_FACTOR 0.75
+
+static double get_prediction_decay_rate(const VP9_COMMON *cm,
+ const FIRSTPASS_STATS *next_frame) {
+ const double sr_decay_rate = get_sr_decay_rate(cm, next_frame);
+ const double zero_motion_factor =
+ (0.95 * pow((next_frame->pcnt_inter - next_frame->pcnt_motion),
+ ZM_POWER_FACTOR));
+
+ return MAX(zero_motion_factor,
+ (sr_decay_rate + ((1.0 - sr_decay_rate) * zero_motion_factor)));
+}
// Function to test for a condition where a complex transition is followed
// by a static section. For example in slide shows where there is a fade
@@ -1172,19 +1193,17 @@
}
}
-// Calculate a baseline boost number for the current frame.
-static double calc_frame_boost(const TWO_PASS *twopass,
+#define BASELINE_ERR_PER_MB 1000.0
+static double calc_frame_boost(VP9_COMP *cpi,
const FIRSTPASS_STATS *this_frame,
- double this_frame_mv_in_out) {
+ double this_frame_mv_in_out,
+ double max_boost) {
double frame_boost;
- // Underlying boost factor is based on inter intra error ratio.
- if (this_frame->intra_error > twopass->gf_intra_err_min)
- frame_boost = (IIFACTOR * this_frame->intra_error /
- DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
- else
- frame_boost = (IIFACTOR * twopass->gf_intra_err_min /
- DOUBLE_DIVIDE_CHECK(this_frame->coded_error));
+ // Underlying boost factor is based on inter error ratio.
+ frame_boost = (BASELINE_ERR_PER_MB * cpi->common.MBs) /
+ DOUBLE_DIVIDE_CHECK(this_frame->coded_error);
+ frame_boost = frame_boost * BOOST_FACTOR;
// Increase boost for frames where new data coming into frame (e.g. zoom out).
// Slightly reduce boost if there is a net balance of motion out of the frame
@@ -1195,7 +1214,7 @@
else
frame_boost += frame_boost * (this_frame_mv_in_out / 2.0);
- return MIN(frame_boost, GF_RMAX);
+ return MIN(frame_boost, max_boost);
}
static int calc_arf_boost(VP9_COMP *cpi, int offset,
@@ -1236,8 +1255,9 @@
? MIN_DECAY_FACTOR : decay_accumulator;
}
- boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
- this_frame_mv_in_out);
+ boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
+ this_frame_mv_in_out,
+ GF_MAX_BOOST);
}
*f_boost = (int)boost_score;
@@ -1274,8 +1294,9 @@
? MIN_DECAY_FACTOR : decay_accumulator;
}
- boost_score += decay_accumulator * calc_frame_boost(twopass, this_frame,
- this_frame_mv_in_out);
+ boost_score += decay_accumulator * calc_frame_boost(cpi, this_frame,
+ this_frame_mv_in_out,
+ GF_MAX_BOOST);
}
*b_boost = (int)boost_score;
@@ -1585,7 +1606,7 @@
gf_group_err -= gf_first_frame_err;
// Motion breakout threshold for loop below depends on image size.
- mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 10.0;
+ mv_ratio_accumulator_thresh = (cpi->common.width + cpi->common.height) / 4.0;
// Work out a maximum interval for the GF group.
// If the image appears almost completely static we can extend beyond this.
@@ -1629,11 +1650,13 @@
if (!flash_detected) {
last_loop_decay_rate = loop_decay_rate;
loop_decay_rate = get_prediction_decay_rate(&cpi->common, &next_frame);
+
decay_accumulator = decay_accumulator * loop_decay_rate;
// Monitor for static sections.
- zero_motion_accumulator = MIN(zero_motion_accumulator,
- get_zero_motion_factor(&next_frame));
+ zero_motion_accumulator =
+ MIN(zero_motion_accumulator,
+ get_zero_motion_factor(&cpi->common, &next_frame));
// Break clause to detect very still sections after motion. For example,
// a static image after a fade or other transition.
@@ -1645,8 +1668,9 @@
}
// Calculate a boost number for this frame.
- boost_score += decay_accumulator * calc_frame_boost(twopass, &next_frame,
- this_frame_mv_in_out);
+ boost_score += decay_accumulator * calc_frame_boost(cpi, &next_frame,
+ this_frame_mv_in_out,
+ GF_MAX_BOOST);
// Break out conditions.
if (
@@ -1655,38 +1679,21 @@
(
// Don't break out with a very short interval.
(i > MIN_GF_INTERVAL) &&
- ((boost_score > 125.0) || (next_frame.pcnt_inter < 0.75)) &&
(!flash_detected) &&
((mv_ratio_accumulator > mv_ratio_accumulator_thresh) ||
(abs_mv_in_out_accumulator > 3.0) ||
(mv_in_out_accumulator < -2.0) ||
- ((boost_score - old_boost_score) < IIFACTOR)))) {
+ ((boost_score - old_boost_score) < BOOST_FACTOR)))) {
boost_score = old_boost_score;
break;
}
*this_frame = next_frame;
-
old_boost_score = boost_score;
}
twopass->gf_zeromotion_pct = (int)(zero_motion_accumulator * 1000.0);
- // Don't allow a gf too near the next kf.
- if ((rc->frames_to_key - i) < MIN_GF_INTERVAL) {
- while (i < (rc->frames_to_key + !rc->next_key_frame_forced)) {
- ++i;
-
- if (EOF == input_stats(twopass, this_frame))
- break;
-
- if (i < rc->frames_to_key) {
- mod_frame_err = calculate_modified_err(twopass, oxcf, this_frame);
- gf_group_err += mod_frame_err;
- }
- }
- }
-
// Set the interval until the next gf.
if (cpi->common.frame_type == KEY_FRAME || rc->source_alt_ref_active)
rc->baseline_gf_interval = i - 1;
@@ -1713,10 +1720,7 @@
// Should we use the alternate reference frame.
if (allow_alt_ref &&
(i < cpi->oxcf.lag_in_frames) &&
- (i >= MIN_GF_INTERVAL) &&
- // For real scene cuts (not forced kfs) don't allow arf very near kf.
- (rc->next_key_frame_forced ||
- (i <= (rc->frames_to_key - MIN_GF_INTERVAL)))) {
+ (i >= MIN_GF_INTERVAL)) {
// Calculate the boost for alt ref.
rc->gfu_boost = calc_arf_boost(cpi, 0, (i - 1), (i - 1), &f_boost,
&b_boost);
@@ -1727,7 +1731,7 @@
(cpi->multi_arf_allowed && (rc->baseline_gf_interval >= 6) &&
(zero_motion_accumulator < 0.995)) ? 1 : 0;
} else {
- rc->gfu_boost = (int)boost_score;
+ rc->gfu_boost = MAX((int)boost_score, 125);
rc->source_alt_ref_pending = 0;
}
@@ -1782,6 +1786,9 @@
}
}
+// TODO(PGW) Re-examine the use of II ration in this code in the light of#
+// changes elsewhere
+#define KF_II_MAX 128.0
static int test_candidate_kf(TWO_PASS *twopass,
const FIRSTPASS_STATS *last_frame,
const FIRSTPASS_STATS *this_frame,
@@ -1811,11 +1818,11 @@
// Examine how well the key frame predicts subsequent frames.
for (i = 0; i < 16; ++i) {
- double next_iiratio = (IIKFACTOR1 * local_next_frame.intra_error /
+ double next_iiratio = (BOOST_FACTOR * local_next_frame.intra_error /
DOUBLE_DIVIDE_CHECK(local_next_frame.coded_error));
- if (next_iiratio > RMAX)
- next_iiratio = RMAX;
+ if (next_iiratio > KF_II_MAX)
+ next_iiratio = KF_II_MAX;
// Cumulative effect of decay in prediction quality.
if (local_next_frame.pcnt_inter > 0.85)
@@ -1870,7 +1877,9 @@
FIRSTPASS_STATS next_frame;
FIRSTPASS_STATS last_frame;
int kf_bits = 0;
+ int loop_decay_counter = 0;
double decay_accumulator = 1.0;
+ double av_decay_accumulator = 0.0;
double zero_motion_accumulator = 1.0;
double boost_score = 0.0;
double kf_mod_err = 0.0;
@@ -2024,42 +2033,38 @@
// Reset the first pass file position.
reset_fpf_position(twopass, start_position);
- // Scan through the kf group collating various stats used to deteermine
+ // Scan through the kf group collating various stats used to determine
// how many bits to spend on it.
decay_accumulator = 1.0;
boost_score = 0.0;
- for (i = 0; i < rc->frames_to_key; ++i) {
+ for (i = 0; i < (rc->frames_to_key - 1); ++i) {
if (EOF == input_stats(twopass, &next_frame))
break;
// Monitor for static sections.
- zero_motion_accumulator =MIN(zero_motion_accumulator,
- get_zero_motion_factor(&next_frame));
+ zero_motion_accumulator =
+ MIN(zero_motion_accumulator,
+ get_zero_motion_factor(&cpi->common, &next_frame));
- // For the first few frames collect data to decide kf boost.
- if (i <= (rc->max_gf_interval * 2)) {
- double r;
- if (next_frame.intra_error > twopass->kf_intra_err_min)
- r = (IIKFACTOR2 * next_frame.intra_error /
- DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
- else
- r = (IIKFACTOR2 * twopass->kf_intra_err_min /
- DOUBLE_DIVIDE_CHECK(next_frame.coded_error));
-
- if (r > RMAX)
- r = RMAX;
+ // Not all frames in the group are necessarily used in calculating boost.
+ if ((i <= rc->max_gf_interval) ||
+ ((i <= (rc->max_gf_interval * 4)) && (decay_accumulator > 0.5))) {
+ const double frame_boost =
+ calc_frame_boost(cpi, this_frame, 0, KF_MAX_BOOST);
// How fast is prediction quality decaying.
if (!detect_flash(twopass, 0)) {
- const double loop_decay_rate = get_prediction_decay_rate(&cpi->common,
- &next_frame);
+ const double loop_decay_rate =
+ get_prediction_decay_rate(&cpi->common, &next_frame);
decay_accumulator *= loop_decay_rate;
decay_accumulator = MAX(decay_accumulator, MIN_DECAY_FACTOR);
+ av_decay_accumulator += decay_accumulator;
+ ++loop_decay_counter;
}
-
- boost_score += (decay_accumulator * r);
+ boost_score += (decay_accumulator * frame_boost);
}
}
+ av_decay_accumulator /= (double)loop_decay_counter;
reset_fpf_position(twopass, start_position);
@@ -2071,14 +2076,12 @@
calculate_section_intra_ratio(start_position, twopass->stats_in_end,
rc->frames_to_key);
+ // Apply various clamps for min and max boost
+ rc->kf_boost = (int)(av_decay_accumulator * boost_score);
+ rc->kf_boost = MAX(rc->kf_boost, (rc->frames_to_key * 3));
+ rc->kf_boost = MAX(rc->kf_boost, MIN_KF_BOOST);
+
// Work out how many bits to allocate for the key frame itself.
- rc->kf_boost = (int)boost_score;
-
- if (rc->kf_boost < (rc->frames_to_key * 3))
- rc->kf_boost = (rc->frames_to_key * 3);
- if (rc->kf_boost < MIN_KF_BOOST)
- rc->kf_boost = MIN_KF_BOOST;
-
kf_bits = calculate_boost_bits((rc->frames_to_key - 1),
rc->kf_boost, twopass->kf_group_bits);
@@ -2152,6 +2155,10 @@
break;
}
if (is_two_pass_svc(cpi)) {
+ if (cpi->svc.temporal_layer_id > 0) {
+ cpi->refresh_last_frame = 0;
+ cpi->refresh_golden_frame = 0;
+ }
if (cpi->svc.layer_context[cpi->svc.spatial_layer_id].gold_ref_idx < 0)
cpi->refresh_golden_frame = 0;
if (cpi->alt_ref_source == NULL)
@@ -2216,11 +2223,6 @@
vp9_clear_system_state();
- if (lc != NULL && twopass->kf_intra_err_min == 0) {
- twopass->kf_intra_err_min = KF_MB_INTRA_MIN * cpi->common.MBs;
- twopass->gf_intra_err_min = GF_MB_INTRA_MIN * cpi->common.MBs;
- }
-
if (cpi->oxcf.rc_mode == VPX_Q) {
twopass->active_worst_quality = cpi->oxcf.cq_level;
} else if (cm->current_video_frame == 0 ||
@@ -2286,6 +2288,18 @@
rc->frames_till_gf_update_due = rc->baseline_gf_interval;
if (lc != NULL)
cpi->refresh_golden_frame = 1;
+
+#if ARF_STATS_OUTPUT
+ {
+ FILE *fpfile;
+ fpfile = fopen("arf.stt", "a");
+ ++arf_count;
+ fprintf(fpfile, "%10d %10d %10d %10ld\n",
+ cm->current_video_frame, rc->kf_boost, arf_count, rc->gfu_boost);
+
+ fclose(fpfile);
+ }
+#endif
}
configure_buffer_updates(cpi);
@@ -2325,6 +2339,7 @@
if (cpi->common.frame_type != KEY_FRAME &&
!vp9_is_upper_layer_key_frame(cpi)) {
twopass->kf_group_bits -= bits_used;
+ twopass->last_kfgroup_zeromotion_pct = twopass->kf_zeromotion_pct;
}
twopass->kf_group_bits = MAX(twopass->kf_group_bits, 0);
diff --git a/vp9/encoder/vp9_firstpass.h b/vp9/encoder/vp9_firstpass.h
index aaa6b03..0b82d32 100644
--- a/vp9/encoder/vp9_firstpass.h
+++ b/vp9/encoder/vp9_firstpass.h
@@ -93,8 +93,6 @@
double modified_error_min;
double modified_error_max;
double modified_error_left;
- double kf_intra_err_min;
- double gf_intra_err_min;
#if CONFIG_FP_MB_STATS
uint8_t *frame_mb_stats_buf;
@@ -110,6 +108,7 @@
int sr_update_lag;
int kf_zeromotion_pct;
+ int last_kfgroup_zeromotion_pct;
int gf_zeromotion_pct;
int active_worst_quality;
diff --git a/vp9/encoder/vp9_ratectrl.c b/vp9/encoder/vp9_ratectrl.c
index 94c0b64..9b6c773 100644
--- a/vp9/encoder/vp9_ratectrl.c
+++ b/vp9/encoder/vp9_ratectrl.c
@@ -126,10 +126,10 @@
int i;
for (i = 0; i < QINDEX_RANGE; i++) {
const double maxq = vp9_convert_qindex_to_q(i, bit_depth);
- kf_low_m[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.125, bit_depth);
- kf_high_m[i] = get_minq_index(maxq, 0.000002, -0.0012, 0.50, bit_depth);
+ kf_low_m[i] = get_minq_index(maxq, 0.000001, -0.0004, 0.150, bit_depth);
+ kf_high_m[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
arfgf_low[i] = get_minq_index(maxq, 0.0000015, -0.0009, 0.30, bit_depth);
- arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.50, bit_depth);
+ arfgf_high[i] = get_minq_index(maxq, 0.0000021, -0.00125, 0.55, bit_depth);
inter[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.90, bit_depth);
rtc[i] = get_minq_index(maxq, 0.00000271, -0.00113, 0.70, bit_depth);
}
@@ -175,7 +175,7 @@
double correction_factor,
vpx_bit_depth_t bit_depth) {
const double q = vp9_convert_qindex_to_q(qindex, bit_depth);
- int enumerator = frame_type == KEY_FRAME ? 3300000 : 2250000;
+ int enumerator = frame_type == KEY_FRAME ? 2700000 : 1800000;
// q based adjustment to baseline enumerator
enumerator += (int)(enumerator * q) >> 12;
@@ -610,7 +610,7 @@
if (frame_is_intra_only(cm)) {
active_best_quality = rc->best_quality;
- // Handle the special case for key frames forced when we have75 reached
+ // Handle the special case for key frames forced when we have reached
// the maximum key frame interval. Here force the Q to a range
// based on the ambient Q to reduce the risk of popping.
if (rc->this_key_frame_forced) {
@@ -880,6 +880,7 @@
return q;
}
+#define STATIC_MOTION_THRESH 95
static int rc_pick_q_and_bounds_two_pass(const VP9_COMP *cpi,
int *bottom_index,
int *top_index) {
@@ -894,16 +895,31 @@
ASSIGN_MINQ_TABLE(cm->bit_depth, inter_minq);
if (frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi)) {
- // Handle the special case for key frames forced when we have75 reached
+ // Handle the special case for key frames forced when we have reached
// the maximum key frame interval. Here force the Q to a range
// based on the ambient Q to reduce the risk of popping.
if (rc->this_key_frame_forced) {
- int qindex = rc->last_boosted_qindex;
- double last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
- int delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
- last_boosted_q * 0.75,
- cm->bit_depth);
- active_best_quality = MAX(qindex + delta_qindex, rc->best_quality);
+ double last_boosted_q;
+ int delta_qindex;
+ int qindex;
+
+ if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+ qindex = MIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+ active_best_quality = qindex;
+ last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+ delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+ last_boosted_q * 1.25,
+ cm->bit_depth);
+ active_worst_quality = MIN(qindex + delta_qindex, active_worst_quality);
+
+ } else {
+ qindex = rc->last_boosted_qindex;
+ last_boosted_q = vp9_convert_qindex_to_q(qindex, cm->bit_depth);
+ delta_qindex = vp9_compute_qdelta(rc, last_boosted_q,
+ last_boosted_q * 0.75,
+ cm->bit_depth);
+ active_best_quality = MAX(qindex + delta_qindex, rc->best_quality);
+ }
} else {
// Not forced keyframe.
double q_adj_factor = 1.0;
@@ -972,18 +988,12 @@
}
}
- // Clip the active best and worst quality values to limits.
- active_best_quality = clamp(active_best_quality,
- rc->best_quality, rc->worst_quality);
- active_worst_quality = clamp(active_worst_quality,
- active_best_quality, rc->worst_quality);
-
- *top_index = active_worst_quality;
- *bottom_index = active_best_quality;
-
#if LIMIT_QRANGE_FOR_ALTREF_AND_KEY
vp9_clear_system_state();
- {
+ // Static forced key frames Q restrictions dealt with elsewhere.
+ if (!((frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi))) ||
+ !rc->this_key_frame_forced ||
+ (cpi->twopass.last_kfgroup_zeromotion_pct < STATIC_MOTION_THRESH)) {
const GF_GROUP *const gf_group = &cpi->twopass.gf_group;
const double rate_factor_deltas[RATE_FACTOR_LEVELS] = {
1.00, // INTER_NORMAL
@@ -997,27 +1007,43 @@
int qdelta = vp9_compute_qdelta_by_rate(&cpi->rc, cm->frame_type,
active_worst_quality, rate_factor,
cm->bit_depth);
- *top_index = active_worst_quality + qdelta;
- *top_index = (*top_index > *bottom_index) ? *top_index : *bottom_index;
+ active_worst_quality = active_worst_quality + qdelta;
+ active_worst_quality = MAX(active_worst_quality, active_best_quality);
}
#endif
+ // Clip the active best and worst quality values to limits.
+ active_best_quality = clamp(active_best_quality,
+ rc->best_quality, rc->worst_quality);
+ active_worst_quality = clamp(active_worst_quality,
+ active_best_quality, rc->worst_quality);
+
if (oxcf->rc_mode == VPX_Q) {
q = active_best_quality;
// Special case code to try and match quality with forced key frames.
- } else if ((cm->frame_type == KEY_FRAME) && rc->this_key_frame_forced) {
- q = rc->last_boosted_qindex;
+ } else if ((frame_is_intra_only(cm) || vp9_is_upper_layer_key_frame(cpi)) &&
+ rc->this_key_frame_forced) {
+ // If static since last kf use better of last boosted and last kf q.
+ if (cpi->twopass.last_kfgroup_zeromotion_pct >= STATIC_MOTION_THRESH) {
+ q = MIN(rc->last_kf_qindex, rc->last_boosted_qindex);
+ } else {
+ q = rc->last_boosted_qindex;
+ }
} else {
q = vp9_rc_regulate_q(cpi, rc->this_frame_target,
active_best_quality, active_worst_quality);
- if (q > *top_index) {
+ if (q > active_worst_quality) {
// Special case when we are targeting the max allowed rate.
if (rc->this_frame_target >= rc->max_frame_bandwidth)
- *top_index = q;
+ active_worst_quality = q;
else
- q = *top_index;
+ q = active_worst_quality;
}
}
+ clamp(q, active_best_quality, active_worst_quality);
+
+ *top_index = active_worst_quality;
+ *bottom_index = active_best_quality;
assert(*top_index <= rc->worst_quality &&
*top_index >= rc->best_quality);
@@ -1161,11 +1187,12 @@
// better than that already stored.
// This is used to help set quality in forced key frames to reduce popping
if ((qindex < rc->last_boosted_qindex) ||
- ((cpi->static_mb_pct < 100) &&
- ((cm->frame_type == KEY_FRAME) || cpi->refresh_alt_ref_frame ||
+ (((cm->frame_type == KEY_FRAME) || cpi->refresh_alt_ref_frame ||
(cpi->refresh_golden_frame && !rc->is_src_frame_alt_ref)))) {
rc->last_boosted_qindex = qindex;
}
+ if (cm->frame_type == KEY_FRAME)
+ rc->last_kf_qindex = qindex;
update_buffer_level(cpi, rc->projected_frame_size);
diff --git a/vp9/encoder/vp9_ratectrl.h b/vp9/encoder/vp9_ratectrl.h
index 2ced8e6..edfb9fc 100644
--- a/vp9/encoder/vp9_ratectrl.h
+++ b/vp9/encoder/vp9_ratectrl.h
@@ -42,6 +42,7 @@
int sb64_target_rate;
int last_q[FRAME_TYPES]; // Separate values for Intra/Inter
int last_boosted_qindex; // Last boosted GF/KF/ARF q
+ int last_kf_qindex; // Q index of the last key frame coded.
int gfu_boost;
int last_boost;
diff --git a/vp9/encoder/vp9_svc_layercontext.c b/vp9/encoder/vp9_svc_layercontext.c
index 7545d87..8c735ab 100644
--- a/vp9/encoder/vp9_svc_layercontext.c
+++ b/vp9/encoder/vp9_svc_layercontext.c
@@ -19,7 +19,7 @@
const VP9EncoderConfig *const oxcf = &cpi->oxcf;
int layer;
int layer_end;
- int alt_ref_idx = svc->number_spatial_layers * svc->number_temporal_layers;
+ int alt_ref_idx = svc->number_spatial_layers;
svc->spatial_layer_id = 0;
svc->temporal_layer_id = 0;
@@ -286,30 +286,17 @@
count >>= 1;
}
- cpi->lst_fb_idx =
- cpi->svc.spatial_layer_id * cpi->svc.number_temporal_layers +
- cpi->svc.temporal_layer_id;
- if (lc->frames_from_key_frame < cpi->svc.number_temporal_layers)
- cpi->ref_frame_flags &= ~VP9_LAST_FLAG;
+ cpi->lst_fb_idx = cpi->svc.spatial_layer_id;
- if (cpi->svc.spatial_layer_id == 0) {
- if (cpi->svc.temporal_layer_id == 0)
- cpi->gld_fb_idx = lc->gold_ref_idx >= 0 ?
- lc->gold_ref_idx : cpi->lst_fb_idx;
- else
- cpi->gld_fb_idx = cpi->lst_fb_idx - 1;
- } else {
- if (cpi->svc.temporal_layer_id == 0)
- cpi->gld_fb_idx = cpi->svc.spatial_layer_id -
- cpi->svc.number_temporal_layers;
- else
- cpi->gld_fb_idx = cpi->lst_fb_idx - 1;
- }
+ if (cpi->svc.spatial_layer_id == 0)
+ cpi->gld_fb_idx = (lc->gold_ref_idx >= 0) ?
+ lc->gold_ref_idx : cpi->lst_fb_idx;
+ else
+ cpi->gld_fb_idx = cpi->svc.spatial_layer_id - 1;
if (lc->current_video_frame_in_layer == 0) {
if (cpi->svc.spatial_layer_id >= 2) {
- cpi->alt_fb_idx =
- cpi->svc.spatial_layer_id - 2 * cpi->svc.number_temporal_layers;
+ cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
} else {
cpi->alt_fb_idx = cpi->lst_fb_idx;
cpi->ref_frame_flags &= (~VP9_LAST_FLAG & ~VP9_ALT_FLAG);
@@ -331,8 +318,7 @@
lc_lower->alt_ref_source != NULL)
cpi->alt_fb_idx = lc_lower->alt_ref_idx;
else if (cpi->svc.spatial_layer_id >= 2)
- cpi->alt_fb_idx =
- cpi->svc.spatial_layer_id - 2 * cpi->svc.number_temporal_layers;
+ cpi->alt_fb_idx = cpi->svc.spatial_layer_id - 2;
else
cpi->alt_fb_idx = cpi->lst_fb_idx;
}
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index ff02666..6a532b2 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -454,12 +454,20 @@
// In spatial svc the scaling factors might be less then 1/2. So we will use
// non-normative scaling.
int frame_used = 0;
+#if CONFIG_VP9_HIGHBITDEPTH
+ vp9_setup_scale_factors_for_frame(&sf,
+ get_frame_new_buffer(cm)->y_crop_width,
+ get_frame_new_buffer(cm)->y_crop_height,
+ get_frame_new_buffer(cm)->y_crop_width,
+ get_frame_new_buffer(cm)->y_crop_height,
+ cm->use_highbitdepth);
+#else
vp9_setup_scale_factors_for_frame(&sf,
get_frame_new_buffer(cm)->y_crop_width,
get_frame_new_buffer(cm)->y_crop_height,
get_frame_new_buffer(cm)->y_crop_width,
get_frame_new_buffer(cm)->y_crop_height);
-
+#endif
for (frame = 0; frame < frames_to_blur; ++frame) {
if (cm->mi_cols * MI_SIZE != frames[frame]->y_width ||
cm->mi_rows * MI_SIZE != frames[frame]->y_height) {
@@ -481,11 +489,20 @@
}
} else {
// ARF is produced at the native frame size and resized when coded.
+#if CONFIG_VP9_HIGHBITDEPTH
+ vp9_setup_scale_factors_for_frame(&sf,
+ frames[0]->y_crop_width,
+ frames[0]->y_crop_height,
+ frames[0]->y_crop_width,
+ frames[0]->y_crop_height,
+ cm->use_highbitdepth);
+#else
vp9_setup_scale_factors_for_frame(&sf,
frames[0]->y_crop_width,
frames[0]->y_crop_height,
frames[0]->y_crop_width,
frames[0]->y_crop_height);
+#endif
}
temporal_filter_iterate_c(cpi, frames, frames_to_blur,
diff --git a/vp9/vp9_common.mk b/vp9/vp9_common.mk
index e88060c..07a3be8 100644
--- a/vp9/vp9_common.mk
+++ b/vp9/vp9_common.mk
@@ -91,6 +91,8 @@
ifeq ($(CONFIG_VP9_HIGHBITDEPTH),yes)
VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_intrapred_sse2.asm
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_subpixel_8t_sse2.asm
+VP9_COMMON_SRCS-$(HAVE_SSE2) += common/x86/vp9_high_subpixel_bilinear_sse2.asm
endif
# common (c)
diff --git a/vp9/vp9_cx_iface.c b/vp9/vp9_cx_iface.c
index 4b24960..0d6ad26 100644
--- a/vp9/vp9_cx_iface.c
+++ b/vp9/vp9_cx_iface.c
@@ -178,8 +178,6 @@
}
#if CONFIG_SPATIAL_SVC
- if (cfg->ss_number_layers * cfg->ts_number_layers > REF_FRAMES)
- ERROR("Too many layers. Maximum 8 layers could be set");
if ((cfg->ss_number_layers > 1 || cfg->ts_number_layers > 1) &&
cfg->g_pass == VPX_RC_LAST_PASS) {
@@ -188,8 +186,7 @@
if (cfg->ss_enable_auto_alt_ref[i])
++alt_ref_sum;
}
- if (alt_ref_sum >
- REF_FRAMES - cfg->ss_number_layers * cfg->ts_number_layers)
+ if (alt_ref_sum > REF_FRAMES - cfg->ss_number_layers)
ERROR("Not enough ref buffers for svc alt ref frames");
if ((cfg->ss_number_layers > 3 ||
cfg->ss_number_layers * cfg->ts_number_layers > 4) &&
